Agents / experiments
AGENTS.md1 starsChanged 58 days agoApache-2.0View on GitHub
Temporary subtrees outside the main quality bar. Each owns its gate; none is in check-all. An experiment consumes the monorepo's packages to answer one empirical question, then gets promoted or deleted. README.md states the policy and how to add one. This file is what an agent needs when working inside an experiment: what isolation means here, and which gate actually runs. - Nothing here is in make check-all. The root gate will stay green while an experiment is broken, so…
In this repository39 other files it gives its agents.
- agent-core/AGENTS.mdAGENTS.md
- .agents/AGENTS.mdAGENTS.md
- AGENTS.mdAGENTS.md
- behavioral-regression/AGENTS.mdAGENTS.md
- claude-foundation/AGENTS.mdAGENTS.md
- config/AGENTS.mdAGENTS.md
- corpora/AGENTS.mdAGENTS.md
- demo/AGENTS.mdAGENTS.md
- docs/AGENTS.mdAGENTS.md
- examples/AGENTS.mdAGENTS.md
- flow-corpus/AGENTS.mdAGENTS.md
- flow-protocol/AGENTS.mdAGENTS.md
- openspec/AGENTS.mdAGENTS.md
- scripts/AGENTS.mdAGENTS.md
- skills/AGENTS.mdAGENTS.md
- tests/AGENTS.mdAGENTS.md
- claude-foundation/CLAUDE.mdCLAUDE.md
- merge-gate-auditorSkill
- update-executive-reportSkill
- update-executive-reportSkill
- architecture-drift-guardSkill
- commonSkill
- corpus-guardianSkill
- dataset-lintSkill
- deploySkill
- e2e-matrix-sentinelSkill
- eval-corpus-forgeSkill
- hierarchical-recursive-brainstormSkill
- model-benchSkill
- openai-judgeSkill
- openspec-implementation-reviewSkill
- openspec-peer-reviewSkill
- openspec-quality-planSkill
- pre-pr-gateSkill
- project-setupSkill
- quality-gateSkill
- refactoring-decomposerSkill
- repo-invariant-reviewSkill
- test-completeness-guardSkill
The file
Files
AGENTS.md — experiments
Temporary subtrees outside the main quality bar. Each owns its gate; none is in
check-all.
An experiment consumes the monorepo's packages to answer one empirical question, then gets
promoted or deleted. README.md states the policy and how to add one. This
file is what an agent needs when working inside an experiment: what isolation means here,
and which gate actually runs.
Map
| Path | Role |
backend-validation/ | Claimed-versus-observed evidence for the eval-backend decision; own pyproject.toml, own gate, ships unsigned |
backend-validation/PROBES.yaml, RUBRIC.md | Live probes, gated behind an explicit human sign-off that stops the phase chain |
trace-analytics/ | Offline SQL sketches over fixture JSONL using the standard library only |
trace-analytics/expected.json | The golden file check_offline.py --update rewrites |
Diagram
flowchart TD
accTitle: how an experiment relates to the repository gate
accDescr: The root check-all target covers the packages only; each experiment owns a separate make check that the root never invokes, and a promoted experiment leaves as a package or an ADR.
ROOT["make check-all<br/>root plus five packages"]
PKG["the packages"]
subgraph EXP["experiments/ (you are here)"]
BV["backend-validation<br/>make -C ... check"]
TA["trace-analytics<br/>make -C ... check"]
SIGN["PROBES.yaml sign-off<br/>stops until a human signs"]
end
OUT["promoted to a package or an ADR,<br/>or deleted"]
ROOT --> PKG
PKG --> BV
PKG --> TA
BV --> SIGN
BV --> OUT
TA --> OUT
classDef here fill:#e8f0fe,stroke:#1a73e8,stroke-width:2px
class EXP hereRules that bite here
- Nothing here is in
make check-all. The root gate will stay green while an experiment is broken, so run the experiment's own gate yourself. That isolation is the point, not an oversight. - The dependency arrow points one way. An experiment may depend on the harness and the packages; the root and the packages must never depend on an experiment. An experiment that something imports is a package, and should be promoted rather than imported.
- Zero writes outside the subtree.
backend-validationhas a PR-scoped isolation check for exactly this, andtrace-analyticsstates it as an invariant. trace-analyticsdeliberately has nopyproject.toml. A manifest there would pull in dependency and coverage obligations the experiment does not have. Do not add one to make tooling happy.- Keep
README.md's experiment table in sync. Bothbackend-validationandtrace-analyticsmust appear there when present on disk; add a row when you introduce an experiment rather than leaving the index to drift. - Live probes are sign-off gated.
backend-validation's preflight phase stops with a distinct exit code until a human sign-off file exists. Do not route around it.
Verify
make -C experiments/trace-analytics check
make -C experiments/backend-validation checkSubagents
| Task in this directory | Agent | Why |
| Confirm nothing in the packages imports an experiment | explorer | Read-only Grep; the one-way dependency rule has no CI check at the root |
| Run one experiment's own gate and isolate the failure | test-runner | Has Bash; each experiment has a different make check and its own floor |
| Review an experiment before promoting it into a package | narrow-critic | Promotion imports unsigned code into the protected quality bar |
See also
| Doc | Read it when |
README.md | You are adding an experiment and need the policy and the exit criteria |
backend-validation/README.md | You are running the backend validation phases or its sign-off chain |
trace-analytics/README.md | You are touching the SQL sketches or regenerating the golden file |
../docs/CHARTER.md | An experiment is edging toward a declared non-goal and needs escalation |
../docs/AGENTS.md | The experiment has served its purpose and its result belongs in an ADR |
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.

