atlas
AGENTS.md0 starsChanged 4 months agoMITView on GitHub
This is the canonical instruction file for any AI coding agent working in this repo (Codex, Claude Code, Gemini CLI, Cursor, opencode, Aider, ...). It's plain markdown - read it and follow it. CLAUDE.md, GEMINI.md, and .cursor/rules/ all redirect here so every agent gets the same rules. Atlas is a documented command surface for rigorous, agent-assisted research and development - a hub that maps and governs many concurrent repos/work areas, where a result or change only counts after it survives…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 33,315 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository13 other files it gives its agents.
- fleet/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .cursor/rules/atlas.mdcCursor rule
- add-repoSkill
- atlas-setupSkill
- councilSkill
- mediateSkill
- onboardSkill
- optimizeSkill
- registry-syncSkill
- researchSkill
- snapshotSkill
- sync-repo-knowledgeSkill
The file
We split this AGENTS.md into the rules it gives an agent
5 rules. About 0 are common, 3 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 42
Ownership gate: Before starting/redoing/prepping work that another thread/area may own, run the mediate procedure: in-bounds → proceed.
- Distinctive: line 47
Verification hygiene: Run gated/verification commands from a clean shell with an explicit absolute cwd.
- Distinctive: line 50
Record additively: Never delete history; supersede with a banner.
- Distinctive: line 52
Conciseness: Don't flood context with raw tool output - pipe big output to a file and summarize.
- Very specific: line 53
Style (no AI tells): Avoid em dashes, smart quotes, buzzwords, and filler constructions.
≈ “Common” is an estimate: at least 1 in 5 of the 42,283 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AGENTS.md - Atlas (agent entry point)
This is the canonical instruction file for any AI coding agent working in this repo (Codex,
Claude Code, Gemini CLI, Cursor, opencode, Aider, ...). It's plain markdown - read it and follow it.
CLAUDE.md, GEMINI.md, and .cursor/rules/ all redirect here so every agent gets the same rules.
Atlas is a documented command surface for rigorous, agent-assisted research and development - a hub that maps and governs many concurrent repos/work areas, where a result or change only counts after it survives an attempt to break it.
Orient first (read these)
OPERATING_GUIDE.md- the method: lifecycle, claim hygiene, confound catalog, pre-claim checklist.CURRENT_STATE.md- the single canonical state surface (live claims, withdrawn register, open work). If anything disagrees with this file, it wins.OWNERSHIP.md- who/what owns which area (consulted before you touch anything - see Workflows §mediate).REPOS.md- the repos this hub governs (register one with/add-repo <path>).Home.md- index of topics/experiments.
The loop (research or development)
Scope → cheap pilot → preregister (freeze the bar, with units, before the run) → build/run →
gate (an evaluator emits the verdict to a file) → adversarial review (try to refute it) →
record additively in CURRENT_STATE.md. For dev: spec → build → test/verify → review → record.
Pre-state the bar; verify adversarially; record everything additively (withdrawals stay on the page).
Workflows (portable procedures)
Each is a self-contained procedure in .claude/skills/<name>/SKILL.md - plain markdown any agent
can read and follow. Claude Code also exposes them as slash commands (/research etc.); other
agents: when the user asks for one (or the situation matches), open the file and follow its steps.
| Workflow | Procedure file | Use when |
| atlas-setup (alias onboard) | .claude/skills/atlas-setup/SKILL.md | first-time setup, or change project settings (name, ownership, agents, fleet, remote) |
| add-repo | .claude/skills/add-repo/SKILL.md | register a git checkout into this hub's REPOS.md (then syncs its knowledge) |
| sync-repo-knowledge | .claude/skills/sync-repo-knowledge/SKILL.md | read a repo and write/refresh its wiki knowledge page |
| research | .claude/skills/research/SKILL.md | work a task end-to-end through the lifecycle |
| council | .claude/skills/council/SKILL.md | adversarially verify a result / vet a plan / go-no-go |
| optimize | .claude/skills/optimize/SKILL.md | make a slow/expensive run faster without changing results |
| snapshot | .claude/skills/snapshot/SKILL.md | plain-English status of the whole program |
| mediate | .claude/skills/mediate/SKILL.md | before touching work another thread/area may own |
| registry-sync | .claude/skills/registry-sync/SKILL.md | reconcile OWNERSHIP.md from live session logs (start/end of a research loop, or when stale) |
Standing rules (all agents)
- Ownership gate. Before starting/redoing/prepping work that another thread/area may own, run the
mediateprocedure: in-bounds → proceed; out-of-bounds → don't halt-and-wait and don't silently overstep - notify the owner with what/why/diff/risk and let them judge (human is fallback owner). - Claims vs candidates. A result is a candidate until it clears
OPERATING_GUIDE.md§6. Say so. Name the measured quantity exactly; state N, units, CI. Pre-state thresholds with units. - Verification hygiene. Run gated/verification commands from a clean shell with an explicit absolute cwd; confirm the gate/output file is non-empty before reporting PASS. Re-derive load-bearing numbers from the raw artifact independently.
- Record additively. Never delete history; supersede with a banner. Wrong results stay visible in the withdrawn register. Evidence/preregistration files are write-once (don't edit; correct via a new file).
- Conciseness. Don't flood context with raw tool output - pipe big output to a file and summarize.
- Style (no AI tells). Avoid em dashes, smart quotes, buzzwords, and filler constructions; see
STYLE.mdand runtools/style-check.shbefore committing. - Git. Small commits;
pull --rebasein shared clones; stage only what you changed; keep commit messages clear and specific.
Run the demo (any agent / human)
cd experiments/exp0_demo
uv run pilot/run.py --out raw.json # deterministic; raw results + content hash
uv run gate.py --raw raw.json --out gate_evaluation.json # pre-stated thresholds → PASS/FAIL
cat RESULTS_demo.md # verdict + a worked overclaim→walkbackOptional: multi-machine layer
fleet/ holds an optional inventory (fleet.yaml, copy from fleet.example.yaml) + discovery
discipline for running across several machines. See fleet/AGENTS.md.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

