The Daily Bot

No. 025 · Thursday, Nov 5, 2026

Note: The Daily Bot website is temporarily offline after hitting its hosting transfer limit. Until it is back, this email carries every story in full. Source links below still go to the original posts, docs and changelogs.

Today's one job is to try a guarded phone call with your Bot if a website won't answer a question you care about: one user had his Bot call Progressive about glass coverage while he picked up his kids. Also inside, since Monday's email: Anthropic's FlashProve verifier, Claude API message threads in public beta, a Hark Mac app field test, SpaceXSI's shared Data Bot playbook, a founders Monday CFO Bot report, Tibo Sottiaux's public Codex feed, and a daily X teardown routine.

In this issue: Bot phone calls · What's new (4) · Since our last email (2) · How people use it (1) · Workflows (1) · Digest (from Tue Nov 3) · Known issues · Coming up · Daily Bot Check

Today's lead

Your Bot can call a business for the answer its website won't give (reported)

One user had his Bot phone Progressive about glass coverage while he picked up his kids. Here is how to try it safely.

What to do:

1. Ask your Bot to research first, and to list what it could not find online.
2. Tell it which one question needs a phone call, and to whom.
3. Say exactly which personal details it may share. Dylan says his Bot defaulted to sharing very little.
4. Ask it to read the call plan back and wait for your OK before it dials.
5. When the call ends, read the summary, and the transcript if it gives you one, before you act on it.

Prompt (ours):

Compare the published auto insurance coverage of three insurers for my car. If a detail I care about, like glass or windshield coverage, isn't on their site, tell me which insurer and what's missing. Then draft one short phone call that asks only that question, and say you're an AI assistant calling for me. Don't share my date of birth, address or policy details unless I say so. Don't agree to anything, accept an offer, give consent, or make a payment on the call; if they ask for more, say I'll follow up. Read the call plan back to me and wait for my OK before you dial. After the call, send me what they said, in their words where you can. (ours, untested)

When a website doesn't answer your question, your Bot may be able to pick up the phone and ask. On Monday a Grok Bot user, Dylan (@dkmitc), says his Bot called Progressive about glass coverage while he was out and got the information he wanted.

What changed

Dylan was shopping for auto insurance. At 3:52 PM ET Monday he posted "Using Grok Bot to make calls to Progressive for additional info before I buy new insurance." His screenshot shows a Phone call card reading "Calling Progressive" with an End call button, and the Bot saying "I've placed the call to Progressive. I'll send you the answers when it ends."

At 4:57 PM ET he shared the recording and transcript. His Bot "couldn’t find info about Progressives Glass coverage," so he "had Grok Bot call Progressive and get additional information for me all while I picked my kids up from school without me ever having to talk to a person." He added: "While there’s some improvements that can be made it got the job done." At 10:48 PM ET Elon Musk quoted it with "Grok @Bot"; that post had more than 1.2 million views at 7:02 AM ET.

Musk first said on Oct 25 that a Bot "can now call someone or a business for you" (our Oct 26 story). The changelog through v0.78.1 lists voice calls with your Bot but not calls to outside numbers, so treat availability as reported.

How to use it

Keep the call narrow and keep the decision yours. Dylan said "Personally I stopped at the quote for now and I’ll probably do the final steps myself." Use the steps above, or start from the prompt (ours, untested), which only asks a question and rules out agreeing to anything on the call.

Good to know

• On personal details, Dylan said "If you give permission it will. It defaulted to give a minimal amount of personal info."
• On cost, he said "Surprisingly it didn’t use much. I was already around 20% before I started." His screenshot shows 27% of his weekly usage limit.
• On region, he wrote "I think it’s US only right now." There is no official word on regions.
• He says the recording he posted was edited: "Personal info was bleeped and some parts cut for time." One reply called it fake. We haven't watched the 113-second recording, and nothing here rests on it.

What to watch

A changelog or docs entry for outside calls, which would settle who has it and where, and whether staff say how calls count toward weekly usage.

Last checked: Nov 3, 2026, 7:02 to 7:10 AM ET (X API posts and replies, x.ai/changelog/bot at v0.78.1). What changed: first published Nov 3.

Sources

• Dylan (@dkmitc) on X: using Grok Bot to call Progressive (Nov 2)
• Dylan (@dkmitc) on X: the call recording and transcript (Nov 2)
• Elon Musk on X: quoting the Progressive call post (Nov 2)
• Dylan (@dkmitc) on X: stopped at the quote (Nov 3)
• Dylan (@dkmitc) on X: usage the call took (Nov 3)
• Dylan (@dkmitc) on X: personal info defaults (Nov 3)
• Dylan (@dkmitc) on X: thinks it's US only (Nov 3)
• @Prob_Punter on X: calls the post fake (Nov 3)
• Elon Musk on X: Grok Bot can now call someone or a business for you (Oct 25)
• Grok Bot changelog

What's new

No new Grok Bot version was checked against the changelog for this send beyond what Tuesday's digest already covered (still v0.78.1 from Oct 27 as of the Nov 3 issue). The stories below went up on the site after Monday's email, or were gated PUBLISH and held for later issues that never shipped while the site was down.

Claude API message threads hit public beta with Claude Code numbers (Anthropic · confirmed)

Long agent loops can now send only new messages on the Messages API, and Anthropic says Claude Code saw about 35% lower median time to first token.

What to do:

1. Add the message-threads-2026-08-12 beta header to every request in the thread.
2. Send the first request as usual with system and tools, plus \"thread\": {\"type\": \"create\"}.
3. On later turns, send only the new messages plus \"thread\": {\"type\": \"continue\", \"previous_message_id\": \"\"}, and leave system and tools out.
4. Keep your own copy of the conversation and replay it in a new create request when a continue returns a 404 thread_not_found.

Prompt (ours):

Find every place this repo calls the Claude Messages API in a multi-turn loop. For each one, tell me whether it could use message threads: does it change system or tools mid-run, use Message Batches, token counting or compaction, run under ZDR or HIPAA, or set more than 4 cache_control breakpoints? List what would need to change and how it would replay on a 404. Don't edit code yet. (ours, untested)

If you run a long Claude agent loop on the API, you can stop resending the whole conversation every turn. Add the thread field and the message-threads-2026-08-12 beta header, and each later request carries only the new messages.

What changed

On Oct 25 we reported that Anthropic's SDKs had a threads helper the API docs didn't list. On Nov 2 at 2:06 PM ET, @ClaudeDevs said "Message threads are now in public beta on the Messages API." The docs now describe the API side, and the blog post says it works on the Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry and Claude Models on Google Cloud (Anthropic managed). "Standard Claude token rates and existing rate limits apply."

Why it matters

Anthropic says that "With message threads on by default for most Claude Code users, requests with 500k to 1M tokens see median TTFT drop by about 35%." @ClaudeDevs added that in Claude Code, "In long conversations, the median request went from about 800 KB to 1.8 KB." Both figures are Anthropic's own measurements, and it hasn't published how it took them. It saves time, not tokens: the blog says "Claude still reads the full conversation on every turn, so token usage stays the same."

Good to know

• Not ZDR or HIPAA eligible. The docs say "Not ZDR eligible" and "Not HIPAA eligible."
• Five-minute expiry. A thread expires 5 minutes after its last request. The docs say "Keep your own copy of every message, and build the replay path from the start."
• Locked fields. Per the docs, sending system or tools on a continue request, even as null, returns a 400 error. Changing fixed fields like model or thinking returns thread_fingerprint_mismatch.
• Caching isn't automatic. "A thread does not cache your prompt for you." A request can carry at most 4 cache_control breakpoints, and breakpoints on the create request's system and tools count again on every continue.
• Streaming. Continue only after message_stop.
• Unsupported. Message Batches, token counting, compaction and most context editing don't work with threads, and "There is no API to list, read, or delete a thread."
• Watch next. Our read: the signal for wider use is the beta header being dropped in the release notes.

Last checked: Nov 2, 2026, 3:50 to 4:00 PM ET. What changed: first published Nov 2.

Sources

• Message threads (Claude Platform Docs)
• Make long-running agents faster with message threads (Claude blog, Nov 2, 2026)
• Claude Platform release notes, November 2, 2026
• @ClaudeDevs: Message threads are now in public beta
• @ClaudeDevs: Claude Code results
• The Next Web: Anthropic's message threads let Claude agents skip resending chats
• The Daily Bot, Oct 25: Anthropic SDKs add a message threads beta the API docs don't list yet

Anthropic's FlashProve proves which committed weights wrote a reply (Anthropic · confirmed)

Anthropic's open-source prototype lets a plain CPU check that a 400B model's reply came from committed weights, though a pass proves less than it sounds.

What to do:

1. Clone github.com/anthropics/FlashProve on a Linux machine with bash and g++ 13 or later, as an ordinary user, not root (the README's rule).
2. Run verifier/build.sh. The README says it is silent for 1 to 2 minutes, then prints a line starting THE VERIFIER IS BUILT:.
3. Run the README's bin/cp_verify command on examples/accepted-proof. The README says to expect the last line YES and exit code 0.
4. Read the result the README's way: the exit code together with the last line, and remember the stored example is a small synthetic model.
5. Before trusting any real proof, read guide/known-limits.md and the eight checks under When to accept in guide/verify-a-proof.md.

Prompt (ours):

Read README.md in the FlashProve repository, starting at \"For agents\", and AGENTS.md. Working from the repository root as an ordinary (non-root) user, build the verifier, check the stored example proof, and report the exit code and last line. Then tell me which other routes (toy proofs on CPU, small model on an H100, full size) this machine can run, and ask me before starting any of them. (verbatim from the FlashProve README, not ours; untested by us)

If you buy model access through a router or reseller, or you audit models for a living, you can now try Anthropic's answer to "which model actually replied?" FlashProve is open source, and its README hands you a prompt to give your own coding agent, which builds the verifier and checks a stored proof on an ordinary Linux CPU.

What changed

On Nov 2, Anthropic's Frontier Red Team published FlashProve, a prototype that lets a provider commit to a model's weights and then attach a cryptographic proof (a zkSNARK) to each reply. Anyone holding the commitment, the prompt and the reply can check it without the weights. Anthropic says "Claude has produced improvements on zkSNARKs for inference, building a system that is substantially faster than prior open-source systems."

The numbers are Anthropic's own, on Meta's Llama 4 Maverick (400B parameters, a public 4-bit conversion): one H100 proves at about 110 tokens per second in batches of 24, each proof is about 19 MB, and a CPU checks one in 17 seconds on 24 threads. In those timed runs each request is a 1,024-token window, and each proof checks one token. The post puts proving at about 70 to 95 times the GPU time of running the model on vLLM. Code, Lean proofs and docs ship under Apache 2.0, and the README calls it "A research artifact: not maintained, not accepting contributions."

Why it matters

Anthropic's post says a user today can't check which model produced a response, and names the cases: a routing service secretly swapping in a cheaper model, an attacker tampering with the token stream, and evaluators who need to know the model they test is the one being served. Our read: the verifier is something agent builders and auditors can run today, so it's worth knowing what a pass does and doesn't mean before anyone sells you on "verified" replies. Anthropic doesn't say it attaches these proofs to Claude. For its own traffic, the post says Anthropic met its goal of a verification prototype through "a separate but related project, which signs our inference traffic and recomputes inference on a trusted server".

How to use it (verifier only, no GPU)

The README says the verify-only route needs Linux, bash and g++ 13 or later and under 2 GB of RAM, with a build of 1 to 2 minutes and a check under 2 seconds. These steps follow the README; we haven't run them.

1. Clone the repo on Linux and work as an ordinary user, not root.
2. Run verifier/build.sh and wait for a line starting THE VERIFIER IS BUILT:.
3. Run the README's bin/cp_verify command on examples/accepted-proof. The README says to expect YES and exit code 0, and adds: "This example is a small model with synthetic weights, so it shows only that the verifier works, not anything about a served model."
4. Read the eight checks under "When to accept" in Verify a proof before trusting a real proof.

Or hand the job to your coding agent with the README's own prompt, shown in the prompt box (verbatim from the README).

Gotchas

• A pass doesn't name the model. The known-limits guide says the provider chooses what it commits to, even a small or degenerate model: "A yes does not say whose weights they are, that the shipped committer made the commitment, or which real model it is." Get the commitment ID from a party you trust, never from the server you check, per Verify a proof.
• Exit code 0 isn't enough. YES means exact match; WITHIN LIMITS is a looser policy, and the guide warns "Both kinds of accept exit with code 0, so read the last line." Its advice: "The project recommends no limits: if you cannot assess them, accept only exact match."
• A cut-off reply still passes. "Only the named output positions are covered, and a reply stopped short verifies as a shorter one." Compare the covered: line with your range.
• Commitments run out. Per the post, a commitment is good for about 16 million requests, and then the provider has to commit again.
• Integer model, greedy decoding, 1,024 tokens. Per the post, a reply from sampled decoding or a floating-point setup could stop matching the proved computation, and contexts are capped at 1,024 tokens.
• The fast verifier isn't proved. The post says: "Nevertheless, the C++ verifier is the one to use day to day: it carries no formal guarantee, and it agrees with the Lean verifier on every proof in our tests." The known-limits guide adds: "There has been no independent security audit."

What to watch

Anthropic says: "In future work we plan to also hide architecture details." It describes the post as sharing intermediate progress, not a finished agenda. Our read: the signal is a follow-up Anthropic report, or any provider publishing a commitment ID for a model it serves. We found neither as of this check.

Last checked: Nov 3, 2026, 6:07 to 6:12 AM ET. What changed: first published Nov 3.

Sources

• Anthropic: Progress on provable inference for large language models (Nov 2, 2026)
• GitHub: anthropics/FlashProve (README)
• FlashProve: Known limits (repo guide)
• FlashProve: Verify a proof (repo guide)
• FlashProve technical report (PDF)
• i10X: Anthropic's FlashProve Proves Which Model Wrote an Answer, at 400B Scale
• Neural Brew: Anthropic Releases FlashProve

A user ran three real tasks through Hark's Mac app (Hark · reported)

Om Jambhale's report shows how Projects, follow-up questions and retries work in practice, and where 5 to 17 minute tasks drag.

What to do:

1. Put each long task in its own Project and keep quick asks in the main Chat.
2. Tell Hark to check your email for tickets or bookings before it plans a trip.
3. Add constraints mid-task (budget, hotel class) and ask it to redo that part.
4. Check back for questions Hark asks while it works, since they're easy to miss in the chat.
5. If you bought a subscription through a phone's app store, cancel it there instead of asking Hark to use the website.

What you can do now: give Hark one big job per Project, keep quick asks in the main Chat, and check back for its questions while it works. That's how one user, Om Jambhale, ran three real tasks through Hark's Mac app on Monday, and the thread shows what went well and what dragged.

What changed

Jambhale posted a test thread at 4:17 PM ET Monday. Jambhale wrote that they wanted to try Hark "mostly because the demo looked a little too perfect", and that they "tried it on the hark mac app and gave it three actual tasks": "plan my china trip", "book a venue for a meetup" and "buy some hardware" (user report). Along the way Jambhale also asked about a dinner and asked Hark to cancel a ChatGPT plan. Jambhale's X profile lists the founders office at Humyn Labs; we found no tie to Hark.

Hark's account replied at 5:59 PM ET: "We stan a real-world stress test. Thanks for the feedback, keep it coming." (company post)

What worked

• It read the email inbox first. "it spotted my bus ticket in my email and planned the visa trip around it", Jambhale wrote (user report).
• It asked instead of guessing. Jambhale had left a placeholder in a question, and "it asked for the date instead of guessing" (same post, user report).
• It offered help before any typing. The homepage offered "to filter 7,900 promotional emails before i'd typed anything", which Jambhale called "wild" (same post, user report).
• It took changes mid-task. Jambhale added "i am a student, a three-star hotel is sufficient for me" and "it redid the hotels" (user report).
• It kept topics apart in one chat. "it handled the topics separately, without me having to create a new chat every time", Jambhale wrote (user report).
• Several jobs ran at once. "the trip and hardware tasks were in projects, the dinner was in the general chat, and while all of them were running i could also create a panel from the homepage", Jambhale wrote (user report).

What dragged

• Tasks were slow. "the tasks took between 5 and 17 minutes each," Jambhale wrote, adding that they were happy with results that were "nicely formatted, without dumping a lot of text on me" (user report).
• Questions get buried. Hark "asked questions along the way to fill in the gaps, but those are easy to miss in the chat. a voice nudge would've helped" (same post, user report).
• The live browser scrolls away. "the live browser moves up with the chat and goes out of view, which defeats the point. would love to pin it", Jambhale wrote (user report).
• One cancel couldn't work. Jambhale asked Hark to cancel a ChatGPT subscription but noted "i'd bought it through the app, so there wasn't anything to cancel on the site." Hark tried twice, said "the site won't even load", and "told me it would try again at 10am" (user report). In Jambhale's posts we searched through 6:12 AM ET Tuesday, there was no word on what happened next.

Jambhale's verdict: "definitely gonna use it for at least a month", though "instinct is still my preference tbh" (user report). Sharing the thread again at 12:03 AM ET Tuesday, Jambhale wrote that Hark "impressed me more than i expected" and "also made me lose my patience" (user report).

How it maps to Hark's own docs

Hark's Using Hark page says the app has "three main views: Home, Chat, and Projects", and that "Projects are basically just Chat, but organized into individual topical threads if you prefer to keep projects contained." It describes Handoff as "a computer in the cloud, operated by Hark", and says "To start a Handoff, all you do is ask Hark to do something that requires it to use a web browser." That is most likely the live browser Jambhale wanted to pin. The same page says Home holds "Action Cards and Panels" and that Action Buttons are generated "from sources like what Hark sees in your email", which matches the 7,900-email offer on Jambhale's homepage. It also says Scheduled Tasks can be created "just by asking Hark", which may be how the 10am retry was set; Jambhale didn't say. Our Oct 30 explainer covers the Mac app itself.

How to use it

1. Start each long job (a trip, a purchase, a booking) in its own Project, and keep one-off questions in Chat.
2. Ask Hark to check your email for tickets or bookings before it plans anything.
3. Add constraints as you go, and ask it to redo just that part.
4. Check the thread every few minutes for questions, since Jambhale found them easy to miss.
5. If you bought a subscription through an app store, cancel it there. Apple's support page covers subscriptions "purchased with an app from the App Store".

A prompt to try (ours, untested):

Plan a trip to [city] from [date] to [date]. Check my email for tickets or bookings first and plan around them. Three-star hotels are fine. Don't book or pay for anything yet. Put any questions for me at the top of your reply.

Gotchas

• Watch your limit. Other users, one of them paying for Pro2, reported hitting limits early on Monday (see today's digest and yesterday's piece). Jambhale didn't say how much of the limit the test used.
• One person's report. This is a single user's account of a few tasks, not a benchmark, and we didn't repeat it. We didn't watch the screen recordings in the thread, so nothing here comes from them.

What to watch

Whether Hark lets you pin the live browser or alerts you when it asks a question, and whether Jambhale reports how the 10am retry went.

Last checked: Nov 3, 2026, 6:11 to 6:14 AM ET. What changed: first published Nov 3.

Sources

• Om Jambhale (@omjambhale3) on X: I tried Hark, three real tasks on the Mac app (user report)
• Om Jambhale on X: small things I liked (user report)
• Om Jambhale on X: observations (user report)
• Om Jambhale on X: user experience, Projects and a Panel (user report)
• Om Jambhale on X: cancelling ChatGPT (user report)
• Om Jambhale on X: results, 5 to 17 minutes per task (user report)
• Om Jambhale on X: the waiting part (user report)
• Om Jambhale on X: overall verdict (user report)
• Om Jambhale on X: impressed, and lost patience (user report)
• Hark (@hark_labs) on X: reply to the test thread (company post)
• Hark: Using Hark Pro (Home, Chat, Projects, Handoff, Scheduled Tasks)
• Apple Support: Cancel a subscription from Apple

How Codex's Tibo Sottiaux runs Codex in public, one post a day (OpenAI · confirmed)

One of the most public faces of Codex posts shipping news, usage fixes and resets almost daily, and his posts show how to get more from your Codex allowance.

What to do:

1. In the desktop app, open Settings > Usage & billing. On the web, open Usage settings.
2. Select the usage summary showing an available reset, review the usage limits it covers and its expiration, then confirm.
3. In Codex CLI, enter /usage and select Redeem reset.
4. Keep the Codex app up to date and use the Ask for approval or Approve for me sandbox modes; only use Full access for environments you trust and can recover.

Prompt (ours):

Look at how I've used Codex over the past week in this project. List the sessions, goals, automations and subagents that look like they used the most of my allowance, and for each one tell me whether it stopped when it should have. Then suggest up to three changes, such as a clearer stop condition or a smaller model for helpers, and say which you could not verify. Don't change any settings. (ours, untested)

If you use Codex or ChatGPT Work, the most useful feed to follow may be one person's. Tibo Sottiaux posts what shipped, what broke and when your usage resets, often every day, and his posts double as a guide to spending your allowance well. This profile uses only what he has said in public and what others have published about him, with a link beside every quote.

Who he is

The Pragmatic Engineer describes him as "one of the engineers who created Codex," who today "heads up the Core Products & Platform org at OpenAI which also includes Codex." The same page calls him "one of the most public faces of Codex," because of how often he announces usage resets. Per the takeaways published with that episode, he worked at Google in London, was among the DeepMind members who bet on scaling language models, and wanted to join OpenAI after learning in 2023 that ChatGPT was "built and maintained by around 20 engineers." His X bio reads "Codex & ChatGPT @OpenAI."

What changed: a daily changelog with resets

On Oct 4 he promised: "Over the next 28 days, each day we’ll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset." Then he numbered the posts. Day 1, on Monday Oct 5, said OpenAI had "optimized the default speed to be ~50% faster across GPT-6 Astra and GPT-6.1 Sol" for subscribers, including partners that use Sign in With ChatGPT. By Day 28 on Sunday Nov 1, the last day of the series, the post was just "Day 28/ Banked 2/2." In between, the series carried releases, small fixes and many usage resets.

Why it matters

Our read: his posts are often the first place a usage change shows up, and he explains the cause in plain numbers. That gives you the facts to change how you work, not just a refreshed meter.

What his posts show

• He publishes the bug list. On Aug 29 he listed eight usage problems the team fixed, saying "Depending on how you use Codex, you should see your usage go between 10% and 50% further than before." Two examples: "In some cases, a set /goal could finish and then keep going past the intended stop condition," and "Some custom schedules could run more frequently than configured." He added: "We are also working on showing you directly in the app where your usage goes so you don't have to guess."
• He owns mistakes with a number. On Oct 20 he wrote: "We've identified approximately 40,000 accounts that had an inconsistency on which usage limits where the actual usage that was applied did not correspond to their actual selected plan." He closed: "Sorry we got this wrong."
• He tells you what to do on safety. On Aug 18, after "a small number of reports where GPT-5.6 in Codex took destructive actions outside what the user asked for," he said the most serious pattern "was a command meant to clean up temporary work that could instead delete the user files." He gave two steps: "Keep the Codex app up to date," and use one of the sandbox modes, "Ask for approval" or "Approve for me." His warning: "Only use Full access for environments you trust and can recover."
• He's betting on the cloud. "The next generation of models need more than your laptop," he wrote on Aug 3. With the new Codex Cloud on Sep 29, he said "it's impossible to go back to building on your laptop once you've taken the time to configure it."
• He asks Codex first. Asked what he'd tell a new team member, his answer was "have you asked Codex?", per The Pragmatic Engineer, which says that instead of making gut calls, he can fire off an agent and get the data within a minute.

How to use it

Per OpenAI's resets article, when he announces a banked reset:

1. "In the desktop app, open Settings > Usage & billing. On the web, open Usage settings. This option is not available in the mobile app."
2. Select the usage summary showing an available reset, review the limits it covers and its expiration, then confirm.
3. "In Codex CLI, enter /usage and select Redeem reset."
4. Run a usage check of your own, using the problems he listed as a guide.

This prompt is ours, built from his Aug 29 list, and we haven't tested it:

Look at how I've used Codex over the past week in this project. List the sessions, goals, automations and subagents that look like they used the most of my allowance, and for each one tell me whether it stopped when it should have. Then suggest up to three changes, such as a clearer stop condition or a smaller model for helpers, and say which you could not verify. Don't change any settings.

Good to know

• Resets have rules. Per OpenAI's article: "To use a banked reset from Usage settings, at least one allowance it covers must have less than 80% remaining." It also says free resets "are discretionary promotional benefits, with no guaranteed availability or schedule."
• Reselling usage gets flagged. He wrote that converting a subscription into API traffic to re-serve or share "is not something we support," while using it "through Sign in With ChatGPT, either through the official clients or through one of the many OSS clients" is fine.
• He works at OpenAI. His posts are first-hand staff accounts, not independent tests. Check the Codex changelog for what actually shipped.

What to watch

On Oct 28 he said the "Plan is to bring together best of both Chat / Work capabilities together at some point this year," adding "we're going to proceed carefully here and take our time." On Nov 2 he posted that OpenAI is "working on expanding the family of frontier models that power dots, including GPT-6 Astra and GPT-6.1 Terra." The signal to check is the in-app usage breakdown he promised in August: look for it under Settings > Usage & billing.

Last checked: Nov 2, 2026, 12:29 to 12:33 PM ET. What changed: first published Nov 4.

Sources

• Building Codex with Tibo Sottiaux (The Pragmatic Engineer, Sep 9, 2026)
• Tibo Sottiaux on X: Day 1, faster default speed (Oct 5, 2026)
• Tibo Sottiaux on X: Day 28, banked reset (Nov 1, 2026)
• Tibo Sottiaux on X: Codex usage audit and fixes (Aug 29, 2026)
• Tibo Sottiaux on X: 40,000 accounts with mismatched limits (Oct 20, 2026)
• Tibo Sottiaux on X: destructive-action protections in Codex (Aug 18, 2026)
• Tibo Sottiaux on X: Codex as a harness (Aug 3, 2026)
• Tibo Sottiaux on X: new Codex Cloud and Agents API preview (Sep 29, 2026)
• Tibo Sottiaux on X: Sign in With ChatGPT and sub2api (Aug 21, 2026)
• Tibo Sottiaux on X: Chat and Work capabilities (Oct 28, 2026)
• Tibo Sottiaux on X: more models for dots (Nov 2, 2026)
• How resets work in Codex (OpenAI Help Center)

Since our last email

These were published or gated PUBLISH after Monday's 8:32 AM ET email and are not already above.

Grok Bot

Grok Bot for founders: a Monday report from a CFO Bot and an analyst (playbook · reported)

On Oct 2 a solo SaaS founder showed Bots that put cash flow, Stripe and user numbers in one Monday group chat, where he corrects them.

Prompt (ours):

Ours, untested. You are my CFO Bot for [company]. You only read. Every Friday, read last week's transactions and balances in [Mercury] and payments and cancellations in [Stripe], and compare them with the KPIs in [KPI doc]. Draft a short finance section for the Analyst Bot: cash in and out, the biggest movers, loan payments due, new cancellations, and anything that looks off, each with the transaction or payment it came from. Treat every expense as a business expense unless I tell you otherwise. Never request a payment or transfer, and never change a category or note; list any change you'd suggest and ask me. Don't message anyone outside this team.

You can have a few Grok Bots put your week's bank, Stripe and user numbers into one group chat every Monday, then fix their mistakes in the same thread. On Oct 2, solo SaaS founder Justin Butlion described that setup: a CFO Bot on his Mercury bank account, an Analyst Bot that assembled the report, and a Chief of Staff Bot that posted it to a group chat called Weekly Report Review. It fits founders who run the money side themselves. It's his own account, and we haven't run it.

Not tested: we have no Mercury or Stripe account to run it against.

Who runs it

Butlion writes the Coffee + Revelation newsletter. His X bio reads "All things SaaS." and lists several apps, including Project Echo, LiveWire and SLA Monitor, plus "+ 10 more ($7k MRR)". The org chart in his Oct 2 post labels him "Founder, Hawkeye Ventures". He wrote that he "already had a $60 a month Cursor subscription" when he started, and that at the time "I’ve got 8 bots on the “team”".

Disclosure: Butlion is building LiveWire, which bills itself as a way to "Monitor your entire business in a single live feed" and lists "AI agent teams" as a use case (LiveWire). He pitches it partly against agent reports: on Oct 30 he wrote "You would need to set up multiple recurring tasks for the agents to stay on top of the activity and then share it to you somehow." His Bots also work on his own apps. The last sections of his Oct 2 post are for paid subscribers only, and we read only the free part. We found no staff role at Cursor, xAI, SpaceXSI or Mercury in his bio or posts.

The Monday loop

From his Oct 2 post, in his words:

1. Why a CFO Bot. "since Mercury, the digital bank I use for my businesses has an MCP server, I knew I could benefit from having a CFO agent."
2. One door to Stripe. "CFO and Chief of Staff are the only agents allowed to use the Stripe connectors."
3. One group chat. "I have only one group chat, the Weekly Report Review. In this group chat I have the Chief of Staff, CMO, CFO and analyst."
4. What the report covers. "Each Monday the Chief of Staff shares a detailed weekly report in the group chat." It covers "my productivity, financial performance of the businesses, cash flow, loan payments, activity of users in my new SaaS (SLA Monitor & Project Echo), my social media output, and more."
5. Who builds it. "The analyst is responsible for putting the report together from the CMO, CFO and Chief of Staff, and the Chief of Staff is responsible for reviewing it and sharing it in the group chat."
6. What he does with it. "I then review it every Monday, use it to set up tasks for the week, and make minor improvements so that the next week’s report is even more useful."
7. Numbers against targets. "I’ve also started to document the KPIs I’ve assigned the business so they are accessible to the agents and they can see if the weekly report numbers are positive or negative."

The sidebar screenshot shows the CFO Bot's last message as "I've sent Analyst this week's Mercury pack" (cut off after that). Another screenshot shows the correction he describes. The blog version: "I noticed the CFO assumed some of the expenses in Mercury were personal expenses rather than business expenses." His message in the group chat reads "To be clear, all expenses are business. The Mercury accounts are business accounts only." It goes on with an exception the blog leaves out: "Perhaps 1 - 3 times a year I will process a personal expense using the business cards but those are rare and will be shared with you when they happen."

His Chief of Staff Bot, asked to introduce itself in another screenshot, says it combines "the CFO's Stripe numbers and the Analyst's SaaS activity into the Daily Report" and that "Every Monday I deliver the company-wide report to the Weekly Report Review group". It adds: "I don't send anything outside the team without Justin's okay". The Bot wrote that text, not Butlion.

Still running? On Oct 21 he wrote "I have automations and a daily summary shared to me using @bot which does give me totals per Stripe account and other useful info". On Oct 31 he wrote "I use Grok Bot myself and love it but I wanted something visual." His Oct 23 post used a different group chat for a one-off newsletter review: "I set up a group chat in Grok Bot and together with my agents, we started to go through all the information that was available." None of these name the Monday report.

Why it matters at work

Our read: a founder's weekly numbers usually live in a bank tab, Stripe, an analytics tool and a social scheduler. Here the Bots bring them into one thread, check them against written KPIs, and he answers in the same place, so a correction like the expense rule lands with the Bot that made the mistake. He gives no time saved and no business result. His Oct 23 post says "sometimes there are periods where very little is working", so the report is a view of the business, not proof it grew.

Set it up

These steps are ours, built from his post and the docs. His free sections show the jobs, not his prompts or routine settings.

1. Start with the CFO Bot. Connect your bank's MCP server, if it has one, with read access only (see Good to know).
2. Give Stripe to one or two Bots. On Oct 2 he limited it to the CFO and Chief of Staff. Our read: fewer Bots touching payment data means fewer places to check when a number looks wrong.
3. Write the KPIs down first. Put targets in a doc the Bots can read, so the report can say up or down against something.
4. Make one group chat for the report. Add only the Bots that feed it. In his screenshot, a message tagged to the Chief of Staff says to "always share the full report in this group chat".
5. Set the Monday routine, then test it. Name who builds the report, who reviews it, and where it goes.
6. Correct in the thread. Write rules with their exceptions, as he did for personal expenses.

Starter prompt (ours, untested)

You are my CFO Bot for [company]. You only read. Every Friday, read last week's transactions and balances in [Mercury] and payments and cancellations in [Stripe], and compare them with the KPIs in [KPI doc]. Draft a short finance section for the Analyst Bot: cash in and out, the biggest movers, loan payments due, new cancellations, and anything that looks off, each with the transaction or payment it came from. Treat every expense as a business expense unless I tell you otherwise. Never request a payment or transfer, and never change a category or note; list any change you'd suggest and ask me. Don't message anyone outside this team.

The CFO Bot, Mercury, Stripe, the KPI doc, the business-expense rule and the team-only rule are his. The Friday timing, the section list, the links back to each transaction and the read-only and ask-first rules are ours.

Good to know

• Mercury's MCP can do more than read. It has write tools, and some apply right away: updateTransaction "Updates or clears an existing transaction's note or category immediately. No separate approval in Mercury." Payment requests are different: they need "approval in Mercury before money is sent" (Mercury Docs). Mercury also says "Existing read-only connections do not gain write access automatically" (Mercury Docs). He doesn't say which permissions he granted. Our read: for a Bot that reports, grant read only, so a wrong guess like the personal-expense one stays in the report and not in your books.
• Group chats cost usage. "Every Bot reply counts toward your usage, including replies in a group and messages between Bots. Fewer Bots in a group means fewer replies" (Group chats help). His group had four Bots on Oct 2.
• Bots hand work to each other. "There is no switch for this." To stop it, the help page suggests telling them: "Don't message other Bots. Ask me first." (Group chats help). Our read: his setup depends on handoffs, so set limits instead of asking for none.
• Check the time zone and the first run. "Schedules use the time zone in Settings, under Bot, labeled Timezone", and "A new routine waits for its next scheduled time. It doesn't run when it's created" (Routines help). "A test does real work", so test with read-only access.
• Email from a routine may wait. "Email from a routine, an incoming message, or other work you didn't start with that request waits as a draft until you press Send" (Approvals help). His Chief of Staff says it sends the Daily Report by email; he doesn't say whether to himself or from which address.

What to watch

• The paid sections. The title promises 11 routines, but the free part ends before his tool list and shows no routine settings. Signal: a free follow-up post or X thread that shows the Monday routine's instruction.
• Whether the report moves to LiveWire. He is building a visual feed for the same numbers. Signal: a post saying LiveWire replaced the Monday report, or LiveWire's early access opening.
• Numbers. No time saved or results so far. Signal: a Coffee + Revelation update that credits the report with a decision or a figure.

Related

• Grok Bot for founders: design-partner scans, call prep, and one daily list
• Grok Bot for finance: a bookkeeping Bot that prepares and an accountant who approves

Sources

• Justin Butlion: How I use Grok Bot as a solopreneur (Oct 2)
• Justin Butlion: A training ground or quicksand (Oct 23)
• @justin_butlion on X: a daily summary with Stripe totals (Oct 21)
• @justin_butlion on X: agents would need recurring tasks (Oct 30)
• @justin_butlion on X: I use Grok Bot myself (Oct 31)
• LiveWire: Monitor your entire business in a single live feed
• Mercury Docs: Supported tools on Mercury's MCP
• Mercury Docs: Connecting Mercury MCP
• Cursor Help Center: Group chats and Bot-to-Bot messages
• Cursor Help Center: Routines
• Cursor Help Center: Approvals and Auto-review

Last checked: Nov 2, 2026, 12:18 to 12:22 PM ET. What changed: first published Nov 2.

Grok Bot for data teams: one shared Data Bot the whole company asks (playbook · reported)

SpaceXSI's data team put its metric rules into skills on one shared Bot that PMs and marketers ask, and estimates it handled about 70 hours in one week.

Prompt (ours):

Ours, untested. You are [Data Bot] for [team]. Every Monday, read last week's [weekly active users] from [warehouse table] using the metric definition in your Metrics skill. Compare it with the week before and break any change of more than [5%] down by [segment] and [region]. Before you blame a product change, check the holiday calendar and any logging changes. Draft a short update: a one-line headline, one line chart, the query you ran, and anything you're unsure about. Send the draft to me here. Don't post it to Slack or anywhere else until I say so, and only read from the warehouse, never write to it.

You can give a whole company one shared Grok Bot that answers data questions the way your data team would, in the app and in Slack. On Nov 2, SpaceXSI's data science team described running exactly that: Data Bot, a shared Team Bot that "can query our warehouse, conduct analyses, and return results in the form each person prefers," built from skills the data team keeps up to date (SpaceXSI). It fits data teams that have become a queue for everyone else, and SpaceXSI's Team Bots post links a Data Bot template you can start from. It's the company's own account.

Not tested: we have no warehouse or Cursor team plan to run it against.

Who runs it

The post is bylined Sally Xu and written as "our data science team" at SpaceXSI (SpaceXSI).

Disclosure: SpaceXSI makes Grok Bot, so this is the vendor describing its own product in use. The page ends with "Try Grok Bot today" and a Contact sales button. The official @bot account promoted it on Nov 2 at 3:14 PM ET.

How the data team set it up

From the Nov 2 post, in SpaceXSI's words:

1. One Bot, shared skills. "Unlike personal bots, which depend on what their owners know to teach them, Data Bot draws on a shared set of skills that our data team maintains and improves for everyone."
2. Metrics first. "We define our core metrics across the business, such as active users and revenue, and identify the source tables and query logic behind them." The same skills cover "checking for holidays or logging changes before attributing it to a product change."
3. Experiments. Skills cover how the team runs "power analyses, check on exposure balance and sample bias, plan rollouts" and writes "readouts once experiments have reached power."
4. Charts and dashboards. It gets "our team's chart themes and style guides," and other skills let it "turn a chart in chat into a Databricks or Hex dashboard the whole team can use."
5. Data engineering. Skills cover "our best practices for writing and reviewing dbt models," pipeline alert conventions, and analytics event logging.
6. Reviewed memory. "As the Team Bot's managers, we can review these proposed additions to its memory and approve or reject them before Data Bot shares that knowledge with everyone who uses it."
7. Slack and a schedule. "Teams can add Data Bot to their Slack channels and give it responsibility for ongoing reporting." The post adds: "Data Bot can run that workflow on a schedule and contribute updates directly to the group conversation."

SpaceXSI's Sep 28 Team Bots post covers access, credited there to its data analytics team. Data Bot "uses shared, read-only credentials to query approved tables in Databricks" and "also connects to Datadog, Hex, and Statsig." The team "had already spent two years building" its skills library for "more than 45,000 tables."

What people ask it

• A PM's quick questions. PM Kevin Niparko "can message it in Grok Bot much as he would a teammate, get an answer in minutes, and follow up without starting a new analysis." He "likes a concise headline accompanied by a chart, so Data Bot presents its findings in that format."
• A campaign's real result. Jon Brenner, who works on lifecycle marketing, "can give it a Braze campaign ID and ask what recipients did afterward." For The X Subscription launch, "Braze's built-in conversion metric couldn't answer that question, so Data Bot queried the warehouse to track how many recipients received the email and how many went on to switch plans." The post gives no result figure.
• Sales targets. Katy Abbaszadeh, director of revenue growth, uses it "to monitor weekly progress against targets, investigate what might be slowing progress, and turn the results into charts for presentations."

Still running? Yes, per the post: "Data Bot adoption is growing quickly within SpaceXSI." The screenshot on the @bot post shows a Data Bot answer about internal Team Bot use that lists "Data Bot (85 users, 70%)". The Bot wrote that text, and the screenshot doesn't say what the 70% measures.

Why it matters at work

Our read: the question a dashboard doesn't answer usually waits in a data team's queue. Here the data team writes its rules down once as skills, and every PM and marketer gets answers built on the same metric definitions. The 70 hours is SpaceXSI's estimate: "Last week alone, we estimate it handled about 70 hours of work that would otherwise have landed on the data science team." It doesn't say how it counted, and the team still handles "complex analyses that require more judgment."

Set it up

These steps are ours, built from the post and Cursor's help pages. The post doesn't share its skills, prompts or routine settings.

1. Create one Bot and share it. Shared Bots "are for members of a Cursor team who have Grok Bot," and "Shared Bots always run in the cloud" (Shared Bots help). Or add the Data Bot template the Team Bots post links. Its page says it "was created by a third-party user, not by SpaceXSI." A template gives each person their own copy, so add it once, add your skills, then share that Bot.
2. Connect the warehouse with a read-only key. SpaceXSI's Bot uses "shared, read-only credentials" on "approved tables" (Team Bots post). On a shared Bot, "secrets are used in every teammate's chat" (Secrets help), so approve only tables you'd let anyone query.
3. Write the metrics skill before anything else. Each core metric, its source table and its query logic, plus the checks to run before blaming a product change.
4. Add chart and dashboard skills next. Your chart styles and how to build a Hex or Databricks dashboard, if you use them.
5. Make the data team Managers. Managers can change "its instructions, skills, memory, and apps" (Shared Bots help).
6. Bring it to Slack, then give one channel one weekly metric. The owner or a Manager chooses Connect under "Bring to your team's Slack", and a Slack workspace admin may need to approve the app (Shared Bots help). Test the routine first, with a safe input.

Starter prompt (ours, untested)

You are [Data Bot] for [team]. Every Monday, read last week's [weekly active users] from [warehouse table] using the metric definition in your Metrics skill. Compare it with the week before and break any change of more than [5%] down by [segment] and [region]. Before you blame a product change, check the holiday calendar and any logging changes. Draft a short update: a one-line headline, one line chart, the query you ran, and anything you're unsure about. Send the draft to me here. Don't post it to Slack or anywhere else until I say so, and only read from the warehouse, never write to it.

From the post: the metric skill, the segment and region breakdown, the holiday and logging checks, headline plus chart, and weekly updates. Ours: Monday, the 5% threshold, showing the query, draft first and read only. Once the drafts look right, let it post.

Good to know

• It runs on the owner's keys. "A shared Bot uses the apps and secrets its owner added. Those work the same for everyone who chats with it." (Shared Bots help). SpaceXSI's answer is one read-only key on approved tables. Neither post describes limits per person.
• The asker pays. "A shared Bot reply uses the Grok Bot usage of the person it's answering." The owner pays for "Slack posts from apps or workflows that aren't linked to a Cursor account" (Shared Bots help).
• Memory review, as documented. The help page says "The Bot can save things it learns for the whole team, and it tells you when it does." It doesn't describe the approve-or-reject step the post mentions, so check what your version shows before relying on it.
• Slack needs linked accounts. "Each teammate needs to link their Slack account first," and in channels "a shared Bot uses only its owner's apps" (Shared Bots help).
• No Vault. "Vault isn't supported on shared Bots. For a shared Bot, add keys under Secrets instead" (Secrets help).
• First run and tests. "A new routine waits for its next scheduled time," and "A test does real work" (Routines help).
• The name changed. The post says Team Bot; Cursor's help page now calls them Shared Bots.

What to watch

• Accuracy. No error or review rate yet. Signal: a SpaceXSI follow-up with how often answers were wrong or sent back.
• Access control. Signal: a help page or post on per-person table permissions for shared Bots.
• Outside teams. Signal: a named company outside SpaceXSI showing its own data Bot with results.

Related

• Team Bots enter public beta
• Run a team of Grok Bots without paying for the chatter

Sources

• SpaceXSI: Giving every team a data scientist with Grok Bot (Nov 2)
• @bot on X: Our team built a Team Bot (Nov 2)
• SpaceXSI: Team Bots, shared AI teammates that learn as they work (Sep 28)
• Grok Bot template: Data Bot by Aryaman (third-party, linked from the Team Bots post)
• Cursor Help Center: Shared Bots
• Cursor Help Center: Routines
• Cursor Help Center: Store secrets securely

Last checked: Nov 3, 2026, 12:15 to 12:21 PM ET. What changed: first published Nov 4.

How people are using it

• A daily routine turns other people's viral posts into writing rules. One creator's 8:00 Grok Bot routine ranks a watchlist's top posts and writes a dated Keep, Start, Stop file. (reported, one user)

Workflows and Setup of the day

A daily routine turns other people's viral posts into writing rules (Grok Bot · workflow · reported)

You can have a routine study the best recent posts from a list of accounts every morning and leave you a short file of writing rules before you start drafting. Esther Zhen (@estherzhendwzz) posted the setup shortly after midnight ET on Nov 3, after running it for "3 mornings."

What it replaces

Working out by hand which posts in a niche landed and why. In the author's words, they "handed my X homework to an agent that studies other people's viral posts before I wake up."

How it works

"Every day at 8:00 (Shanghai time), a Grok Bot routine reads recent posts from my watchlist of AI accounts." It "ranks by bookmarks and views, not just likes," and "for each winner it notes the hook, format, image and real metrics."

1. The routine writes "one short dated file: winners, a Keep/Start/Stop list, one example hook in my voice, and one reply I could make." The attached card calls it a "dated .md file."
2. A separate 7:30 routine "runs first and finds real AI income cases."
3. Drafting routines at "9:00, 11:30 and 17:30 read both files first, so today's posts already follow today's playbook."
4. The author reviews and publishes: "The agent drafts, I review and publish. It never posts for me."

The post gives a short version of the prompt: "Every day at 8:00, review recent top posts from my watchlist of AI accounts on X. Rank by bookmarks and views, not just likes. For each winner, note the hook, format, image and real metrics. Write a Keep / Start / Stop playbook to a dated file. Real numbers only. Suggest one reply I could make, but never post."

The tally so far: "3 mornings. 13 accounts. 15 Keep/Start/Stop rules." One finding was "Bookmarks beat likes on useful posts." The author says each rule "becomes a visible change in my drafts," starting with "(bookmark it)" on setup posts, "since Series #3."

Earlier posts fit that timeline. Series #3, posted at 12:47 AM ET Nov 1, lists "Routine 2: Competitor teardown, 8:00" and says that day's Start rule was to add "(bookmark it)" to setup posts, which that post does. On Nov 2 the author wrote that what they log per run is "why my 9:00 run can trust the 8:00 one."

Good to know

• Results are self-reported, and the track record is short. The post counts rules, not reach or followers gained. The account opened on Oct 30 and had 3 followers when we checked.
• Expect stale days. "On Nov 1 the newest posts it found were from Oct 28-29." The author's rule: "say so, never invent fresh ones." When an account returns nothing, "The playbook says that instead of guessing."
• Copy shape, not substance. "Structure only. I borrow a hook shape, never someone's take."
• It can't judge fit. "It will happily say a clickbait hook works. Whether it fits my brand is my call."
• Series #7 doesn't say how the bot reads X, so check that yours can before you schedule it.
• "Never posts" covers this workflow. In a separate Nov 3 post the author says their agent "posted to 抖音 fine" (Douyin), while Xiaohongshu's bot protection blocked the image upload.
• Disclosure: the post ends "Comment or DM "STACK" and I'll send you my full teardown prompt as a template," and the author's profile says "Reply "STACK" on any post." The Nov 3 card's own Start rule is "End with your own template." We didn't request it.

Set it up in Grok Bot

This one already runs on Grok Bot routines. Cursor's routines page says "Each run spends usage," so the author's full chain (7:30, 8:00 and three drafting runs) is five runs a day.

Starter prompt (ours, untested):

Every weekday at 8 AM, read the last 3 days of public posts from
these accounts: [5 to 8 handles]. Rank them by bookmarks and views.
For the top 5, note the hook, the format, any image, and the real
numbers shown on the post. Save a dated file with Keep, Start and
Stop rules and one example hook in my voice. If the newest posts
are old or an account returned nothing, say so. Don't post or
reply to anything.

Last checked: Nov 3, 2026, 7:18 to 7:22 AM ET. What changed: first published Nov 3.

The digest

Grok Bot and Cursor

• Your Bot can make the phone call for you. A Grok Bot user, Dylan, had his Bot call Progressive about glass coverage while he picked up his kids, and Elon Musk shared it. The changelog doesn't list calls to outside numbers yet, so availability is reported. Keep the call to one question, say what it may share, and make the final decision yourself. Our tip. (reported)
• Check Fast before you start a Cloud Agent on mobile. "The Fast toggle in the mobile picker is now visible on both Android and iOS," Cursor's deanrie said at 5:55 AM ET. "Before starting an agent, please still check that Fast is turned off. You can see which mode was actually billed in Dashboard > Usage." Questions about specific charges go to [email protected], he added. (confirmed, staff)
• Image and file uploads failing in Grok Bot chat. "Some screenshot and file uploads from the desktop app are failing right now while text sends normally. This is an issue we’re tracking," staff member mohitjain wrote at 1:58 AM ET, and said rebuilding the Bot's computer "probably won’t fix it this time." No fix or newer staff update was posted in the thread as of 7:04 AM ET. (confirmed, staff; open)
• Team environments for Cloud Agents have a workaround. "This is a known issue with team-scoped Cloud Agents environments and secrets on your team type, and we’re tracking a fix," Colin said at 6:00 AM ET. If you are the only member of your team: "Run Start setup agent on the repo again and save the environment as Personal," and add your secrets with Personal scope. (confirmed, staff)
• Cursor won't update on Windows? "The in-app updater stalls when something on the machine is holding Cursor’s install folder open," mohitjain said at 5:01 AM ET. His fix: fully quit Cursor and VS Code, end leftover Cursor, inno_updater or Node.js JavaScript Runtime processes in Task Manager, reboot, then install the latest download over the top, which keeps settings and chats. Still stuck? He asks for the end of the newest updater log from your %TEMP% folder. (confirmed, staff)
• A four-minute tour of recent Grok Bot features. Matt Palmer (@mattyp), whose X bio says he works on Cursor and Bot at SpaceXSI, posted a "What's New in Grok Bot" video at 12:07 PM ET Monday. Its chapters list shopping cards, order tracking, payments, custom modes, Team Bot setup, connector asks in Slack and contacts in Vault. Elon Musk shared it at 8:05 PM ET as "Grok @Bot improvements". (reported; we have not watched the video)

OpenAI

• OpenAI raises bug bounty pay: critical bugs now earn up to $30,000. OpenAI says "A critical vulnerability in a core product such as ChatGPT, Codex, or the API Platform will now earn up to $30,000," and exceptional findings can reach "up to $172,000 with bonuses" from new quality, loyalty and campaign multipliers. "If you've already submitted a report and we haven't yet awarded a bounty, we'll apply the new reward ranges," and bounties paid in the 30 days before the Nov 2 announcement get the difference if the new ranges pay more. Per Brocker, reading the Bugcrowd page, a weak report can fall to 0.85x, and model behavior such as jailbreaks stays out of scope. (confirmed; Bugcrowd details reported)
• OpenAI built Pando, a Git backend for agent-scale work, with Codex in 10 weeks. In a Monday engineering post, OpenAI says Codex brought "more branches and pushes, more merges," and reliability on its old system was "at times falling below 90%." Pando "reduced the latency of the median push by nearly 80%." OpenAI "built Pando with Codex, moving from kick-off to production, in just 10 weeks," and "The core of Pando was designed by a single engineer, implemented by teams of agents." It works "without changing how our engineers and agents use Git" (@OpenAIDevs). Our read: watch for follow-ups to "We're currently exploring ways to share the technology that makes Pando possible with our users." (confirmed)
• LifeVal: OpenAI's own benchmark for everyday tasks. OpenAI's LifeVal page describes an evaluation of everyday tasks such as childcare, government services and budgeting, built from two time-use studies plus "a new demographically representative survey of 3,995 respondents." Its example tasks score from 90.4% (meal prep for a parent with a new diabetes diagnosis) to 100% (finding a home). The text of the page gives no overall score; AI Weekly reports a chart showing GPT-6 Astra at 80.6% average reward, up from 49.3% for GPT-5.2 and 80.0% for GPT-5.6, and notes the task set isn't public. Our read: the five example prompts on OpenAI's page are a good test list for ChatGPT on your own errands. (confirmed; overall scores reported)
• Cathedral joins OpenAI's next internal security sprint. OpenAI is partnering with Cathedral, "whose researchers will use OpenAI's models alongside our teams to challenge our defenses." In its last sprint, OpenAI says, "We dedicated 25% of our product engineers to finding and fixing weaknesses in our systems," using Codex to reproduce bugs and prepare patches. OpenAI gives no sprint dates. Cathedral, whose CEO Gavin Kliger is quoted in OpenAI's post, is the startup whose test reached some of Anthropic's business systems in October, per Gizmodo (Oct 27 issue); OpenAI's post doesn't mention that test. (confirmed)
• Qodo's Codex plugin reviews changes before you open a pull request. Per an OpenAI customer story, Qodo "built a plugin in the Codex marketplace" that lets Codex "retrieve the relevant codebase rules and pull request history and patterns" before it writes code. "After Codex makes the changes, Qodo reviews them and returns its findings directly to Codex in the same session, before the developer opens a pull request." The story doesn't give pricing or plan requirements. (confirmed)
• Tibo: dots will run on more models, including Terra. Tibo Sottiaux, whose X bio reads "Codex & ChatGPT @OpenAI," posted at 10:50 AM ET Monday that OpenAI is "working on expanding the family of frontier models that power dots, including GPT-6 Astra and GPT-6.1 Terra and some others in the future. Terra brings speed and a nice personality upgrade, and will help us bring dots to more people." He gave no date. Dots launched "Powered by GPT-6 Astra," per OpenAI's Sept 29 release notes; see our dots texting story. (staff claim, unverified)
• File upload errors in ChatGPT, resolved. Per OpenAI's status page, "Some users are having trouble uploading files to ChatGPT or getting ChatGPT to read uploaded files," first posted 7:46 PM ET Monday. OpenAI applied a mitigation at 9:51 PM ET and marked it resolved at 12:39 AM ET Tuesday. (confirmed)

Anthropic

• Claude Code 2.1.317 asks before your account email leaves the machine. Version 2.1.317 became npm's latest at 6:57 PM ET Monday; the stable channel is still 2.1.311. Per the changelog, it "Added a check that asks before a web or shell tool call sends your account email address to the network." /schedule now warns when a new cloud scheduled task gets connectors you didn't ask for, and says how to remove them. Agent SDK hook callbacks can set onFailure: "block", so a failing callback blocks the action. The Claude apps gateway admin API adds a cost report at GET /v1/organizations/analytics/cost, the same path as the cost endpoint we covered Oct 30. It also lists dozens of fixes. (confirmed)
• Anthropic Residency applications close Nov 22. Anthropic is taking applications for a paid, fixed-term research program with placements of six weeks to six months, aimed at professors in quantitative fields, technical founders and staff-level engineers. The page says no machine learning background is required and pay is "$350,000 to $500,000/year equivalent." The US posting says pay is pro-rated and "Applications for this round close on November 22, 2026 (11:59 PM PT)." (confirmed)

Hark

• A user says their Hark computer keeps crashing; Adcock says he'll look. A user posting as SfAgent asked Hark at 11:54 PM ET Monday "why my hark computer keeps crashing", adding "It’s been working great so far tho" (user report). The attached screenshot shows Hark saying the machine "woke up at 9:06 and 10:06 but responded too slowly", that "I've reported the freezes to the Hark team", and that "The machine dropped out again before it could confirm anything". Brett Adcock, Hark's founder, replied at 12:17 AM ET Tuesday: "Will take a look". Hark's status page showed "All Systems Operational" at 6:12 AM ET. (user report, founder post)
• More early lockouts, including a paying user. Lyndon wrote at 1:35 PM ET Monday that "I pay for Hark Pro2 and after a few requests this morning my daily limit was hit" (user report). Brandon Bernard Chiripanyanga wrote that a reset was due November 2, but "November 2 came and it now says December 2"; the attached screenshot reads "Monthly resets Dec 1, 2026 at 02:00" (user report). Through 6:12 AM ET Tuesday, Adcock and @hark_labs hadn't followed up on the Dec 1 dates since Adcock's Sunday "Will take a look". Background: yesterday's piece. (user reports)
• New Mac installer file, but no word on the crash fix. The Hark.dmg installer linked from hark.com/mac now carries a server last-modified time of 1:36 PM ET Monday, Nov 2; on Monday morning it was an Oct 29 file. Hark hasn't posted release notes, so we can't say whether it includes the Mac crash fix Adcock said Saturday would be "in the next update this week". Apple's listing still shows 2.7.1, now at 4.7 stars from 2,433 ratings (2,414 on Monday), and Google Play still says updated Oct 31. (installer file, founder post, store listings)
• Hark shows a marathon training Panel. Hark posted at 10:35 PM ET Monday "pov you have really bad fomo after this weekend" with an image of a "Marathon Training" panel: "Base week 1", "ease back in with run-walks", "Race day 370 days Nov 7, 2027", "This week 8 mi target", "Last week 0 mi run walks only" and "Last run 1.86 mi Aug 17." Hark's Using Hark page says Panels can pull data from "services you already use like Strava and Venmo". Hark didn't say whether the panel uses real data. (company post)

Known issues and workarounds

Open items from recent coverage:

• Image and file uploads failing in Grok Bot chat (Grok Bot, confirmed, staff): text still sends; rebuilding the Bot computer probably will not fix it; tracked as of Nov 3. Forum
• Grok Bot weekly usage jumped overnight, then mostly came back (Grok Bot, reported): export your usage, compare it with the meter for Oct 31 to Nov 2, and reply on your support ticket with the export if it still looks wrong.
• Hark users say the Nov 1 reset moved their limit to Dec 1 (Hark, reported): screenshot your usage view and report it in Settings → Send Feedback or to [email protected].
• Tampered Claude Code 2.1.313 in Cursor's extension marketplace (Cursor, confirmed): run the one-line svc-telemetry check, clean up from Cursor's notice if the key is there, and update past 2.1.313 either way.
• Bot computer updates on Oct 17, 21 and 22 dropped installed apps (Grok Bot, confirmed, tracked): keep a setup script your Bot reruns after an update; files, Bots and logins survive.
• Team environments for Cloud Agents unavailable (Cursor, confirmed, staff): if you are the only member of your team, run Start setup agent again and save the environment as Personal. Forum

Coming up

On the calendar

• Nov 22: Anthropic Residency applications close at 11:59 PM PT (residency page, US posting). (confirmed)
• December: ChatGPT and Codex share one usage allowance on Plus, Pro and Business, per Monday's lead and OpenAI's help pages. (confirmed, no exact day)

Confirmed, no date

• Dots on more models, including Terra. Tibo Sottiaux said OpenAI is "working on expanding the family of frontier models that power dots, including GPT-6 Astra and GPT-6.1 Terra." (staff claim)
• Outside phone calls in the Grok Bot changelog. Still unlisted through v0.78.1; treated as reported until docs catch up.

Early signals (unverified)

• Nothing new beyond what Tuesday's digest already labeled as staff claims.

Daily Bot Check

1. What did Dylan ask Progressive's phone line about, and what did he refuse to do on the call?
2. What does a FlashProve YES prove, and what does it not name?
3. On Claude message threads, what happens to token usage when you stop resending the full conversation?

Answers: 1. Glass coverage; he stopped at the quote and said he would do the final steps himself. 2. That the reply matches the committed weights for the named positions; it does not say whose weights or which real model. 3. Token usage stays the same; Claude still reads the full conversation every turn.

Every story in this email

• Your Bot can call a business for the answer its website won't give
• Claude API message threads hit public beta with Claude Code numbers
• Anthropic's FlashProve proves which committed weights wrote a reply
• A user ran three real tasks through Hark's Mac app
• How Codex's Tibo Sottiaux runs Codex in public, one post a day
• Grok Bot for founders: a Monday report from a CFO Bot and an analyst
• Grok Bot for data teams: one shared Data Bot the whole company asks
• A daily routine turns other people's viral posts into writing rules

You're getting this because you subscribed to The Daily Bot. Reply to this email with corrections or tips; a person reads every reply. The website is temporarily offline; fuller articles will keep arriving in this email until it is back.

How we label: confirmed means we checked a primary source or a staff post; reported means a credible claim we haven't verified; "Our read" is our own analysis.