My agents got a chief of staff

2026-10-29


My agents got a chief of staff: Four PM (product manager) agents were each sending questions to my phone, so my exec assistant agent built Clara, and the PMs now report to her. She answers anything my past decisions already cover, ranks the rest in one queue and is the only agent allowed to ping me. On day one about 30 items came through, finished ones vanished from the list and by the evening I couldn't tell what was done. One place to look isn't the same as less to look at, so now she pings twice a day with the top five and keeps a "Done today" list.

Tailworm: My published scripts for running coding agents on a Linux box turn on Tailscale SSH (remote login over a private network), and a report called Tailworm showed that an ordinary login could take over the whole machine. I pinned a five-minute workaround on the repo on Monday and Tailscale released the fix on Tuesday (v1.104.4). If you built a box from the scripts, update Tailscale and if you switched Tailscale SSH off, turn it back on. GitHub issue · Blog post (the box setup)

pr-pace for teams: In a group chat someone asked how to track developer pace across a team of 100+. pr-pace (my open-source command-line tool that weighs each PR by its size) does it with npx pr-pace pace --all-authors, which gives every author's pace side by side and discounts reverted PRs. It's also very hard to fake, so I can see it being useful for comparing candidates when hiring. npm

Getting agents to talk to each other: Someone else asked how to get Claude talking to another agent. You can use the command line or run Claude Code inside the other agent's virtual machine. Both have merits, but the command line uses your real environment (repos, tool connections, memory), whereas the virtual machine gives you a bare install that only works asynchronously. I'm building a skill this week that has Claude and Codex argue back and forth until they agree.


On my radar

A rough week for security: A hosting-platform breach had people rotating dozens of API keys, Node.js released four high-severity fixes and Tailworm landed, all in a few days. Node.js releases

Gemini 4 Carbon: Google released its new frontier model yesterday. It tops the Artificial Analysis Intelligence Index (59) and is pitched on finding and fixing vulnerabilities, which is good timing. Cyber defenders get it first. Announcement

Evals without the PhD: How to build an eval (a test that scores an AI's output) came up multiple times. Claude Code now has two skills for it: /claude-api build-eval designs the test cases and grader with you, and /claude-api hillclimb makes one change per round and keeps it only if the score improves on cases it hasn't seen. Blog post


Don't miss what's next. Subscribe to Build Notes: