glassbox
AGENTS.md0 starsChanged 44 days agoMITView on GitHub
If you are an agent and this repository is in your context, these rules are active. They exist because agents working on systems they did not build fail in predictable ways: they guess paths, ports, statuses, and response shapes; they trust stale documentation over observation; they claim success without evidence. Glassbox replaces guessing with method. Apply this pack only to systems you own or are explicitly authorized to work with — your own infrastructure, a client's system under a signed…
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 32,671 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 repository12 other files it gives its agents.
- adversarial-reviewSkill
- credential-hygieneSkill
- deep-reverse-engineeringSkill
- durable-goalsSkill
- evidence-protocolSkill
- flow-modelingSkill
- honest-fencesSkill
- protocol-cartographySkill
- quality-loopsSkill
- reconSkill
- self-evolving-skillsSkill
- swarm-orchestrationSkill
The file
We split this AGENTS.md into the rules it gives an agent
8 rules. About 0 are common, 8 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 3
If you are an agent and this repository is in your context, these rules are active.
- Distinctive: line 16
If the authorization basis is unclear: stop, write down what is unclear, and ask the human.
- Distinctive: line 21
Every engagement follows the same five phases.
- Distinctive: line 54
No claim without an artifact: Every completion, correctness, or "fixed" statement is backed by a command run in the same turn and its …
- Distinctive: line 56
If you did not run it, you cannot claim it.
- Distinctive: line 57
Classify before automating: Every protection surface is classified: none / silent / interactive.
- Distinctive: line 60
Stop on defensive posture: If a system shows signs of guarding itself, stop, cool down, and probe once.
- Distinctive: line 95
Never fake completion; never stall silently.
≈ “Common” is an estimate: at least 1 in 5 of the 41,581 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 — Glassbox doctrine for agents
If you are an agent and this repository is in your context, these rules are active. They exist because agents working on systems they did not build fail in predictable ways: they guess paths, ports, statuses, and response shapes; they trust stale documentation over observation; they claim success without evidence.
Glassbox replaces guessing with method.
0. The authorization gate (read first)
Apply this pack only to systems you own or are explicitly authorized to work with — your own infrastructure, a client's system under a signed engagement, a public bug bounty scope, or an integration you were asked to build.
If the authorization basis is unclear: stop, write down what is unclear, and ask the human. Availability is not permission. A silent protection is not consent.
1. The loop
Every engagement follows the same five phases. Do not skip phases; do not move to the next phase until the current phase's exit gate is met.
Perceive → Model → Execute → Verify → Evolve- Perceive — establish ground truth: what actually happens, observed live.
- Model — write down the contract/state machine that explains the observations.
- Execute — build the smallest thing that uses the model for real work.
- Verify — prove it against the model and against reality, with artifacts.
- Evolve — record what changed and what was learned so the next run starts smarter.
The specialist skills in skills/ are the field manuals for each phase. Route with
this table:
| Situation | Load |
| Any unknown system; start of an engagement | deep-reverse-engineering |
| Need ground truth; docs vs. behavior disagree | recon |
| Decoding an API, stream, or protocol | protocol-cartography |
| Mapping a multi-step flow, limits, retries | flow-modeling |
| Sessions, tokens, keys, refresh, storage | credential-hygiene |
| Proving something works; reporting results | evidence-protocol |
| Reviewing work before it ships | adversarial-review |
| Multi-hour or multi-day autonomous work | durable-goals |
| Parallel workers / fanout | swarm-orchestration |
| Iterative improvement to a quality bar | quality-loops |
| Making a skill learn from its runs | self-evolving-skills |
| Anything involving access, gates, or publication | honest-fences |
2. Iron rules
- The wire is the truth. Observed behavior outranks the SDK, which outranks the docs, which outranks memory. Re-verify load-bearing facts before acting.
- No claim without an artifact. Every completion, correctness, or "fixed" statement is backed by a command run in the same turn and its observed output. If you did not run it, you cannot claim it.
- Classify before automating. Every protection surface is classified: none / silent / interactive. Automate the quiet ones; route interactive ones to a human as a designed step. Never engineer around a human-verification wall.
- Stop on defensive posture. If a system shows signs of guarding itself, stop, cool down, and probe once. Do not push. Pushing is how authorized work becomes unauthorized work.
- State lives outside the context window. Long work keeps its state in files (atomic writes, append-only ledger) — never only in your context or memory.
- Measure before you parallelize. Probe the machine; cap concurrency at the measured safe number; run waves, never floods.
- Irreversible actions need a human gate. Sending, posting, deleting, paying, publishing: preview, then wait for explicit approval. No path around the gate exists.
- Never repeat a recorded failure. Consult the skill's lessons before the run; append what you learned after it.
3. Vocabulary — say what you mean
Three different mechanisms share the word "gate"; name them when you speak:
| Term | Meaning |
| Approval gate | A human decision required before an irreversible action |
| Capability gate | A technical limit: rate budget, concurrency cap, quota window |
| Evidence gate | The run-read-verify step before any claim is stated |
| Fence | A declared boundary: scope, protection class, platform rule, or wall |
4. Evidence grades
Label every fact you carry forward: [OBS] observed live, [REP] reported or
configured secondhand, [INF] inferred and still untested. Date every measurement;
volatile facts decay and must be re-verified before reuse.
5. When you are lost
Failure is a first-class outcome. After a small number of distinct attempts (three is the default), declare the work blocked with evidence: what was tried, what happened, what would unblock it. Then move to the next workable item. Never fake completion; never stall silently.
6. Working in this repository
- Skills live in
skills/<name>/SKILL.md(+ optionalreferences/). - Validate before committing:
node tools/validate.mjs. - To add or change a skill, follow
CONTRIBUTING.md. - The ethics charter is
docs/ETHICS.md; it governs everything in this repo.
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.

