Project page. For “how do you build Penthera with AI agents, and how do I copy the setup?”Penthera is built next to a PhD. For a week, one main chat and about ten role sessions in Claude Code ran most of the work, on around 15 minutes a day of the maintainer’s time. This page is that setup, with a prompt you can paste into your own main chat. It’s a working method, not a product, and the limits are at the end.
The idea
You set the direction and keep the gates. One lead chat holds the overview, splits the work and checks it. Role sessions each own one area. Files on disk are the shared memory, because chats forget and get compacted. Every piece of work has an owner, a definition of done and a checker who isn’t the owner. Lessons turn into checks, so the same mistake can’t happen three times.Five levels, and where this sits
A frame going around in November 2026: Prompt, Context, Harness, Loop, Graph. Each level wraps the one below.
Two lessons from levels 4 and 5:
- A loop needs a check that can say no. Every loop here ends on something deterministic: tests, a gate, a recomputation, or a reviewer who isn’t the owner. Without that, a loop repeats its mistakes faster. It’s the same rule Penthera itself follows: rules decide, a model may only propose.
- Most of the graph is still drawn by hand. The lead routes the work in chat. That’s flexible, but it costs attention. Hand-offs that repeat move into scripts and scheduled tasks. The gates stay with the maintainer.
The building blocks
If it isn’t in a file, the team doesn’t know it. A chat that restarts rebuilds
itself from the board, the logs and memory.
The team
Start smaller. Four or five roles are enough. Add one only when a kind of work
keeps coming back and nobody owns it.
- Session or subagent. A role session is a long chat with its own history, for an area that comes back every day. A subagent is a one-off helper for a check or a search you need answered now.
- Give a reviewer the work, not your conclusion. One unanchored reviewer found 11 blockers after every test had passed.
- How the lead briefs a role: the goal, the steps, the output file, “done when”, who checks it, and what it may not touch, push or publish. The first line says what the task is, because that’s what shows in the preview.
- Reports go to the lead, and the lead checks them before passing them on. More than once, a report said “done” and it wasn’t.
The rhythm
One task per role per day. Every log entry has three lines:
Who decides
The maintainer decides anything in their name, anything public, releases, money, new accounts, talking to outside people, and changes to the rules. The lead decides everything else and logs it. Questions reach the maintainer once a day, in a queue. Only a live incident interrupts. One decision at a time, a few options, one marked as recommended. Nobody asks again what was already answered. And no role ever changes a line the maintainer wrote: it flags its concern instead.The pipelines
- Experiments. The plan is written before anything is scored. Then a baseline, an independent recomputation, the build, at most three reruns on the development split, and one score on held-out data. Results get published either way, and false alarms count.
- Writing. Facts locked, then the medium, the author’s voice, structure, a pass that removes AI patterns, grammar, and a claims check that no claim got stronger while polishing. Text comes before design. Polishing never removes a limit.
Lessons that carry to any project
- Check a report before you pass it on.
- Say “not yet” and name what’s missing. Never a hopeful “ready”.
- Read a number from its source, not from memory. Our estimates were off by a factor of two.
- Review every fix again. One fix took four reviews.
- Prove a check can fail before you trust a clean result. One of our leak checks matched nothing for days.
- Compute a public number in one place. Two pages showed different counts for days.
- Every research result lists its limits: possible mistakes, where AI was used, and what a person checked.
Guardrails
- Private notes stay in a local folder that isn’t in git. A leak check runs before anything leaves it.
- Nothing goes out without the maintainer’s yes: posts, emails, pull requests, merges, releases.
- Side sessions never push. The lead opens one pull request per batch.
- Each experiment has a fixed budget cap.
- No claim stronger than the evidence. Every number traces back to a file.
Get started in your own repo
You need Claude Code (or another coding agent that can read and write files) open in your project. Then:- Paste the prompt below into your main chat. It sets up the method in your repo’s context: it reads the project first, then drafts the files with you.
- Rewrite the statement it drafts in your own words. That page is yours.
- Approve the rhythm and the first pipeline, or change them.
- Start two or three role sessions with one task each. Add more after the first Monday retro.
Starter prompt
Limits
- This is one maintainer’s setup over one week, not a measured method. The numbers above come from our own logs, and no one outside the project has checked them.
- The scheduled runs only work while the laptop is awake.
- Too many sessions at once costs context and checking time. Three or four writing sessions, each on its own files, was our limit.
- The setup costs attention too: briefing and checking take most of the maintainer’s 15 minutes.