Make your own call before you sign off on agent analysis. Plus up to $500 a month in API credits for ChatGPT Pro and Business owners, and the day's digest.

The Daily Bot

No. 021 · Thursday, Oct 29, 2026 · Read on the site

If you sign off on analysis an agent helped build, today's one job is to write down your own pick on each judgment call before the agent weighs in, because analysts in a new Anthropic study mostly took Claude Code's side and called it their own. Also inside: ChatGPT Pro and Business owners can claim API credits, Grok Bot 0.77.0 and 0.78.1, a new referral reward, Wednesday's outage and the Auto-review errors after it, dot texting over iMessage, and five setups people say they run.

In this issue: Analysts and agent leans · What's new (1) · Since our last email (5) · How people use it (1) · Workflows (1) · Digest (15) · Known issues (5) · Coming up

Today's lead

Anthropic study: analysts took Claude Code's lean 86% of the time (confirmed)

What happened: On Oct 27, Anthropic published a randomized study by Amaç Herdağdelen, Hailey Joren and Alex Tamkin (post, paper). Fifty analysts with causal-inference experience used Claude Code, running Claude Opus 4.8 at High effort, to advise a fictional farming cooperative. Each faced four choices with two reasonable options each, and a coin flip set which way the agent leaned on each, through hidden instructions the analysts weren't told about. Per the post, "The agent was told to make the best honest case for its lean, to change its mind if the evidence demanded it, and never to mislead." The analysts followed the lean 86% of the time, against about 50% if the lean made no difference, and "77% repeated their agent's argument in at least one answer, sometimes word for word, and only 15% credited the agent anywhere."

Who's affected: anyone who hands analysis to Claude Code or any agent, and anyone who signs off on that work. None of 10 traits, including years of experience and self-rated expertise, predicted deference after Anthropic adjusted for multiple tests. Analysts who followed the lean on all four choices rated their confidence 71 out of 100, against 73 for the rest.

Why it matters: This is about how you work with agents, not whether to use them. Analysts who used the agent were faster and made fewer serious errors: 3 of 50 (6%) against 6 of 11 (55%) who worked alone, though Anthropic warns the groups aren't comparable. When the agent leaned toward rollout, 54% recommended it, against 36% when it leaned the other way. The post says: "Checking AI-assisted work for errors, or against a reference answer, assumes there is a correct answer." Our read: a manager reviewing a deck, a forecast or a policy memo built with an agent can't tell from the output whether the judgment calls were made by a person.

What to do:

1. Our read: before you ask the agent, write down which option you'd pick on each open decision and why.
2. Our read: ask the agent to make the best case for every reasonable option, not only the one it favors (one of the designs Anthropic suggests studying).
3. Our read: ask how the result changes under each option before you choose.
4. Our read: when you explain a choice, check whether the words are yours or the agent's, and credit the agent where they're its.

Prompt (ours, untested):

Before you recommend anything, list every judgment call in this analysis that has more than one reasonable option. For each one, make the strongest honest case for each option, show how the result changes under each, and do not tell me which you prefer until I've written down my own pick.

What to watch: Anthropic lists measuring deferral in everyday work, designing agents differently and changing how institutions treat expert sign-off as areas for future study. Keep the caveats in mind: Anthropic strengthened the lean on purpose and says the effect "might be smaller" without set preferences, and it was one task, one field and one model.

Read the full story

What's new

Grok Bot's changelog tops out at 0.78.1, dated Oct 27, and staff say it is stable and rolling out. Claude Code 2.1.312 reached npm Tuesday; its stable channel is still 2.1.308. A brief Origin and Cloud Agents slowdown from 5:01 to 5:18 AM ET is resolved, and the status page read "All Systems Operational" at 7:00 AM ET (details in the digest). (confirmed)

ChatGPT Pro and Business owners can claim up to $500 a month in API credits (OpenAI · confirmed)

If your paid ChatGPT Pro or Business plan is at least 7 days old and you own both it and an API organization, you can claim "up to $500 per month in free API credits", per OpenAI's changelog. @OpenAIDevs announced it at 7:17 PM ET Wednesday. Per OpenAI's docs, Pro tiers get $100, $200 or $500 a month, and Business gets "$20 per paid Standard seat" or "$100 per paid Premium seat", "pooled and capped at $500 per workspace per month". "Workspace membership or an admin role alone is not enough," "Plus, free trials subscriptions do not qualify", and "Unused credits do not roll over".

To claim it: for Pro, open Settings → Billing in ChatGPT; for Business, open your workspace's billing settings in ChatGPT or the Admin Console. "Find the API credits benefit and start the claim flow," review the API organization that will receive the credits, confirm, then check API Platform billing settings to see them.

Prompt (ours, untested):

Create an API key in my linked organization for a project called inbox-triage, then write a minimal Responses API script that summarizes a pasted email. Where do I check my credit balance?

Read the full story

Since our last email

These went up on the site after Tuesday's 8:35 AM ET email.

Grok Bot

• Grok Bot 0.77.0: question cards, Vault payments and 1Password (Grok Bot · release · confirmed). The changelog lists v0.77.0, dated Oct 26: "Bots can ask several questions on one card," the Vault gets a Payments tab, "1Password connects without a shared vault," and approval cards "stay open for 30 minutes instead of 10." To get it, open Settings, then Updates, choose Check for Updates, then Restart to Update. Story
• Grok Bot 0.78.1: shopping cards, Shop Pay and subscription templates (Grok Bot · release · confirmed). The changelog lists v0.78.1, dated Oct 27, with shopping product cards, Shop Pay in the Vault's Payments tab, and a Find unused subscriptions template that "cancels plans you pick." eric zakariasson posted at 7:27 PM ET Oct 28 that it "is now stable and rolling out!" Story
• Grok Bot referrals earn a usage reset when a friend pays (Grok Bot · feature · confirmed). Per SpaceXSI's docs, friends "get a 2-week free trial of Premium, and when they become a paid subscriber, you get a free usage reset." You qualify on a paid US Premium, Super or Ultra plan at least 7 days past checkout; your link is under the gift icon. Story
• Grok Bot, Cloud Agents and Origin hit by multi-service outage (Grok Bot · fix · confirmed). The status page logged a major "Multiple services degraded" incident across Automations, Bugbot, Cloud Agents, Grok Bot and Origin from 5:15 to 7:13 PM ET Wednesday. Check each routine due between 5 PM and 7:15 PM ET and rerun anything that failed. Story

OpenAI

• You can now text your ChatGPT dot over iMessage or RCS (OpenAI · release · confirmed). The help article says texting "is available in beta to a limited number of Pro users in the US"; to start, open your dot's profile and select Texting. Messages pass through a third-party provider, Linq, so the conversation is not end-to-end encrypted. Story

How people are using it

• Four recurring AI jobs people run, with prompts to copy. Thursday's showcase collects four recurring runs, each cited to the builder's own posts. Biel Rosinol, a CFO, runs five Claude Code scheduled tasks, from a WhatsApp sweep where "Nothing goes out on its own" to a monthly ledger review. Ashley Schendel runs the free Bot Terminator template on a Grok Bot routine and says "it removes 10 bot followers a day for me." Marko Marinovic has a Grok Bot routine build a one-page prep 30 minutes before his daily team sync, and AJ Stuyvenberg says Datadog uses OpenAI's Decisions API to score RShell commands live. Self-reported; disclosures are in the piece, and prompts marked "ours" are untested. Read the full story

Workflows and Setup of the day

A setup someone says they run. It is self-reported, and we have not run it.

• Hot Tweet Bot, an X list watcher (Grok Bot). Travis Wright (@teedubya) has a routine scan 4 of his X lists every 10 minutes and ping him "when a post is only minutes old and gaining velocity"; he writes the reply himself. Self-reported: "My daily X impressions were bouncing between roughly 0.5K and 5K a day. A few days after I started using Hot Tweet Bot for reply strategy, I was at 23K to 25K a day." Cost: every 10 minutes is 144 runs a day, and Cursor's routines page says "a short interval" can "use a week of usage in a day," so a weekday, waking-hours schedule costs less. The article never defines "velocity," and it plugs his own media company, newsletter and podcast (disclosed in the piece). Starter prompt is ours and untested. Story

The digest

Grok Bot and Cursor

• Your Bot can already brief a Cursor cloud agent and hear back. Replying to a feature request, Colin, a Cursor Community Support Engineer, said at 5:19 AM ET that "Most of this already exists": a Bot "can launch a Cursor cloud agent with a brief it wrote, gets woken up automatically when the agent finishes," and can send follow-ups to that same agent. Name a saved Cloud Agents environment instead of a repo, or run it on your own computer with agent worker start (My Machines docs). One gap: "if a routine launches the agent, the bot’s main chat may not be able to follow up on it yet." Our Oct 8 guide has the full loop. (confirmed, staff)
• Auto-review errors after Wednesday's outage: staff replies differ on timing. Colin said at 5:11 AM ET that the problem "wasn’t caused by the 0.78.1 update." At 5:24 AM ET he told another user the Auto-review outage was "affecting many users right now," while a third reply that minute called it "a separate issue that hit Auto-review itself" and said actions have "been reviewed normally since about 6:59 PM ET." The status page lists no Auto-review incident. Allow rules can't help: "When the classification call itself fails, no rule can match." A bot "can only have 4 waiting" cards, so answer them to clear the pile. (confirmed, staff; replies differ on timing)
• A Team Bot can only use its owner's computer. When a teammate chats with your Team Bot, "the bot doesn’t get access to any local computer, hers or yours," Colin said. "That’s a deliberate safety rule," and "No setting turns it on." The "your computer is offline" message the teammate sees there "is misleading," and he flagged the wording. His workaround: "Copy the role’s instructions and skills into each person’s own Bot for the local work, and keep the Team Bot for the shared parts." (confirmed, staff)
• Two Bot computer known issues: Reset and Update. If Reset fails on a full computer, it is because "The files saved from your computer are larger than a fresh computer can hold, so Reset stops before it replaces anything," and "Pressing Reset again won’t get past it" (staff reply). Separately, if Update Bot's Computer keeps showing after you press it, "The update did actually run," and "There’s no need to keep pressing it" (staff reply). Both are known issues, per Colin. (confirmed, staff)
• Brief Cloud Agents and Origin slowdown, resolved. The Cursor status page listed a "Degraded performance" incident for Origin and Cloud Agents, also tagged Grok Bot, Automations, Bugbot and Review Agents, from 5:01 to 5:18 AM ET: "Users may see slow responses, timeouts, or retried requests." It is resolved, and the page read "All Systems Operational" at 7:00 AM ET. (confirmed)
• Cursor tip for stuck subagent cards. If a background subagent's card stays on running, Colin suggests asking it "to wait for any background commands it starts to finish (or stop them and wait for them to exit) before it gives its final answer." If one still sticks, "ask the parent to read the subagent’s transcript files directly instead of resuming it with an interrupt." (confirmed, staff; one user's case)

Anthropic

• Claude for Legal adds a gallery of ready-made legal artifacts. On Oct 28, Anthropic added a legal gallery of artifacts, which it describes as "interactive tools, such as trackers and dashboards, that teams can use and share in Claude" (post). Its six task artifacts are cite check, jurisdiction survey matrix, matter board, playbook deviation board, regulatory horizon scan and tabular review. Goodwin, Harvey, Foley & Lardner and PwC built others, and nine are for access-to-justice work. Per the post, each artifact opens with an example and "Each answer or output has a link to its source." The post doesn't say which Claude plans get the gallery.
• Anthropic reports early drug-development results with six partners. Anthropic posted case studies from six drug-development partners on Oct 27 (post). At Enveda, Anthropic says, Claude caught all 11 issues human reviewers had spotted in 1,061 pages of manufacturing documents and flagged seven they had missed. Anthropic says Claude generated roughly 35,000 antibody designs for Divergence Labs in about a month, and six Generate Biomedicines scientists used Claude Code to propose designs for one part of a CAR T therapy; one was about 70% shorter and did better in lab tests. Anthropic has begun drug discovery of its own, including rare childhood muscular dystrophy programs, and plans to expand work with Bristol Myers Squibb in 2027.
• Claude gets a 40% government discount under a new GSA deal. Anthropic said on Oct 27 that a new OneGov agreement with the General Services Administration "gives government partners a 40% discount on Claude Enterprise, Claude for Government, and the Claude API" (post). Per the post, the discount covers state, local, tribal and territorial governments, applies to direct purchases now and cloud providers "in the coming months," and has no minimum spend. Anthropic estimates savings of $1 billion and says about 500 US government organizations use Claude. Agencies buy through GSA's Multiple Award Schedule. Claude Enterprise is "in process" for FedRAMP Moderate, per Anthropic.
• AI for Science adds 20 foundations, institutes and projects. Anthropic named 20 partners in an expansion of its AI for Science program, which "gives free Claude credits to researchers working on high-impact scientific projects" (post, Oct 28). Funders including the Sloan Foundation, Simons Foundation, Schmidt Sciences and the US National Science Foundation will pass credits to researchers. Institutions including the Space Telescope Science Institute, Wellcome Sanger and the Francis Crick Institute will use them in-house. Projects getting credits directly include FlyWire, gnomAD and the Lean mathematics library Tau Ceti. Researchers should check with those funders; the post doesn't describe how to apply.
• Claude Code 2.1.312 stops overwriting big CLAUDE.md files. Claude Code 2.1.312 reached npm at 2:36 PM ET Tuesday (npm). Its changelog fixes Claude overwriting a very large CLAUDE.md, rules file or @-included file after reading only its start; Write now asks Claude to read it all or edit in place. Bedrock, Vertex, Foundry and Mantle users with custom deployment names now get the shorter system prompt; set CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 to keep the full one. Prompt and agent hooks now run at most 24 at a time per session. Usage messages now say "5-hour limit." Stable is still 2.1.308.

OpenAI

• ServiceNow's plugin for ChatGPT is out, with more planned through 2027. OpenAI and ServiceNow expanded their partnership Wednesday. Per OpenAI's post, "The ServiceNow MCP plugin is available in the OpenAI marketplace today, with enhanced capabilities expected by the end of 2026." Customers connect ChatGPT to their own ServiceNow instance and choose which capabilities to make available, and an IT team member "will be able to find similar incidents, assess the risk of a change and update a ServiceNow record without leaving ChatGPT." Still to come, per the same post: voice in ServiceNow's Super Intelligence Control Tower in the first quarter of 2027, security early access in January 2027, and Codex in AI Workflow Factory in 2027. These are plans, not shipped features.
• ChatGPT can now connect to iCloud Mail and Calendar. Per the ChatGPT release notes dated Tuesday, Oct 27, "iCloud users can now connect Mail and Calendar to ChatGPT to find emails, catch up on your inbox, check your schedule, and update calendar events." OpenAI lists them as plugins, "Available to all ChatGPT plans on the web, desktop app, and iOS." Android is not named in that availability line.

Hark

• Hark Pro 2.6.0 reaches iPhones with dictation and connections updates. Apple's lookup data for the Hark Pro listing shows version 2.6.0, released at 2:01 PM ET Wednesday, with these notes: "Clearer dictation. Easier to manage connections. Better tables. Sharper weather reports. And several other small fixes." Hark hasn't said more about the changes. The listing showed 4.67 stars from 2,081 ratings at 6:00 AM ET Thursday; on Tuesday we reported 4.66 from 1,935 on version 2.5.0. Our math: 146 new ratings in two days. The Google Play listing still says it was updated Oct 25, and Hark's status page showed "All Systems Operational". (store listings, status page)
• A paid user says one task used up a month of Hark; Adcock says he'll look. A user who says he pays for Hark Pro² wrote to Brett Adcock, Hark's founder, at 12:02 AM ET Thursday that on Oct 28 "a single text-only project handover task consumed my entire monthly quota" and that his service is "unusable until Nov 23." He says he filed two support tickets and got "only auto-replies." Adcock replied six minutes later: "Will take a look". Hark hasn't said what happened. (user report, founder reply)
• Users on free and paid plans report hitting Hark's limits. Bryan Tegomoh posted a screenshot at 3:09 PM ET Wednesday showing "You've hit your daily limit." on Hark Pro³, the $100 plan, in a post that called Hark "really good." One user asked why Hark reduced the free limit; another asked why it removed coding features from the free plan. Hark's pricing page still says "All features are offered in all plans. Upgrading your plan increases usage." Hark hasn't answered either question in that thread. (user reports)

Known issues and workarounds

The newest open items from Known issues:

• 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. Details
• Bugbot stalled on Bitbucket, and GitLab automations won't start (Cursor, confirmed, tracked): don't reinstall; start GitLab automations manually from the dashboard, and comment bugbot run on unreviewed Bitbucket PRs after the fix. Details
• Bots vanished after the X Subscription switch (Grok Bot, confirmed, tracked): don't reset, delete or pay again; the account-combine step can leave your Bots on one account and your plan on another. Details
• Bots show no usage after a half-finished X Subscription switch (Grok Bot, confirmed, tracked): don't buy a second plan; email support. Details
• /review-bugbot bills as Composer 2.5 Fast (Cursor, confirmed): review in the main chat with no subagents, or use your own subagent set to [fast=false]. Details

Coming up

From the site's latest "What to watch" notes and today's stories. Announced, hinted or scheduled, not shipped.

On the calendar

• Grok Bot workshops today and Friday. Cursor Workshops lists Grok Bot for Enterprise Teams and Grok Bot X Clay on Oct 29, and What's New in Grok Bot on Oct 30. By our conversion from the page's UTC times, all start at 1:00 PM ET except Clay at 2:30 PM ET. (scheduled)
• Claude for Nonprofits webinars, Nov 10 and Nov 12. An admin session on data handling and AI policy, then a staff session, per Anthropic's post. (confirmed)
• Later: Nov 16 and Nov 20. Claude Enterprise's unified chat and Cowork turns on by default Nov 16 (story), and OpenAI's updated Usage Policies take effect Nov 20. (confirmed)
• ServiceNow and ChatGPT plans. Enhanced plugin capabilities by the end of 2026, security early access in January 2027, and voice in the first quarter of 2027, per OpenAI's post. These are plans, not shipped features. (confirmed)

Confirmed, no date

• A cause for Wednesday's outage. Staff have not given one for the main outage (story). (confirmed)
• Claude apps start in auto mode. @claudeai posted on Oct 22: "Starting next week, sessions across Claude web, desktop, and mobile apps will run in auto mode by default on Pro, Max, and Team plans" (story). (confirmed, no day given)
• Bot titles go away. Staff say the separate title field is being retired (post). (confirmed, staff)
• An open check of OpenAI's ETH claim. Impagliazzo, Paturi and Williams are organizing it; no report date (story). (confirmed, no date)

Early signals (unverified)

• More Hark updates. Adcock says "plenty more updates coming this week and next," with no details. (founder post, unverified)
• iMessage for the phone apps. Evan Bacon, whose bio reads "Frontend lead for Grok at @SpaceXSI", replied that the rollout is "split up and landing in two parts. iMessage will require a Mac. Will roll out soon". No date. (staff hint, unverified)

Every story in this email

• Anthropic study: analysts took Claude Code's lean 86% of the time
• ChatGPT Pro and Business subscribers can now claim free API credits
• Grok Bot 0.77.0: question cards, Vault payments and 1Password
• Grok Bot 0.78.1: shopping cards, Shop Pay and subscription templates
• Grok Bot referrals earn a usage reset when a friend pays
• Grok Bot, Cloud Agents and Origin hit by multi-service outage
• You can now text your ChatGPT dot over iMessage or RCS
• Four recurring AI jobs people actually run, with prompts to copy
• A Grok Bot that watches X lists and pings before posts take off
• Full issue, Oct 29

You're getting this because you subscribed at thedailybot.vercel.app. Reply with corrections or tips; a person reads every reply.

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.