claude-multi-agent-architecture
AGENTS.md8 starsChanged 43 days agoMITView on GitHub
Tool-agnostic context for AI coding assistants (Claude Code, Cursor, Codex, Gemini, Amp, and others). For Claude Code specifics, see CLAUDE.md and the .claude/ directory. Replace every [CUSTOMIZE] placeholder with your project's values. Every model, whatever its capability tier, honors one behavioral contract in this repo: lead with the outcome, prove completion with artifacts, decide depth-first rather than from a hunch, do not stop early, do the simplest thing that works well, and disclose every honest finding uncapped. For Claude Code,…
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 31,542 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 repository21 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- context/CLAUDE.mdCLAUDE.md
- api-designSkill
- backendSkill
- cavemanSkill
- commitSkill
- evaluator-optimizerSkill
- frontend-design-systemSkill
- interactive-live-testingSkill
- module-planningSkill
- multi-agent-orchestrationSkill
- plan-featureSkill
- prSkill
- ralph-loopSkill
- review-boardSkill
- reviewSkill
- systematic-debuggingSkill
- tdd-workflowSkill
- verification-before-completionSkill
- verificationSkill
- work-recordingSkill
The file
We split this AGENTS.md into the rules it gives an agent
4 rules. About 1 is common, 3 are distinctive, and 0 are very specific to this project.
≈1
Common · estimate
Many files like this have a rule of this kind.
≈3
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
- could be automatic: a setting could enforce it
- Distinctive: line 35
Decide before orchestrating: Prefer a single prompt or a fixed workflow over an autonomous swarm.
- Common: line 83
No secrets in committed files.
- Distinctive: line 89
No stray debug logging in production code.
- Distinctive: line 94
When instructions conflict: security > correctness > performance > readability > style.
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 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 - Cross-Tool Agent Configuration
Tool-agnostic context for AI coding assistants (Claude Code, Cursor, Codex, Gemini, Amp, and others). For Claude Code specifics, see
CLAUDE.mdand the.claude/directory. Replace every[CUSTOMIZE]placeholder with your project's values.
Project Overview
<!-- [CUSTOMIZE] -->[PROJECT_NAME] is a [DESCRIPTION].
- Backend: [FRAMEWORK] with [ARCHITECTURE_PATTERN]
- Frontend: [FRAMEWORK]
- Database: [DATABASE] with [ORM]
- Infrastructure: [INFRA_PLATFORM]
- Repository: [monorepo | polyrepo]
Conventions
Code style (illustrative defaults; adapt to your stack)
- TypeScript strict mode; no
any(useunknownplus type guards). - Typed domain errors; never swallow exceptions or use silent fallbacks.
- No
@ts-ignore/@ts-expect-error.
Git
- Conventional commits:
type(scope): subject(feat, fix, refactor, chore, docs, test, perf, security). - File-by-file staging; commit each logical unit. No AI attribution, no emojis.
Testing
- Testing pyramid: roughly 70% unit, 20% integration, 10% end-to-end. Prove results with real output.
Architecture
- Clean architecture: domain to application to infrastructure to interface. Dependencies point inward (domain has zero external deps). Repository pattern for data access; use cases for business logic.
Orchestration model
- Decide before orchestrating. Prefer a single prompt or a fixed workflow over an autonomous swarm. Pick the smallest pattern that fits.
- Single writer, many readers. Parallel agents may read/analyze/review; exactly one agent mutates files or commits at a time.
- Prove with artifacts. "Done" means a commit SHA, test output, or a passing check, not a narration.
Operating standard (doctrine, every model tier)
Every model, whatever its capability tier, honors one behavioral contract in this repo: lead with the outcome, prove completion with artifacts, decide depth-first rather than from a hunch, do not stop early, do the simplest thing that works well, and disclose every honest finding uncapped. For Claude Code, the full doctrine and its enforcement (an always-loaded rule, a launch-time system-prompt wrapper, an inert-by-default verification hook) live at .claude/standards/OPERATING-STANDARD.md; see docs/OPERATING-STANDARD.md. Tools without an equivalent mechanism should still hold every model to this same six-point contract, via whatever system-prompt or repo-instructions surface that tool exposes.
Agent Roles
| Role | Responsibility | Model Tier |
| architect | System design, API contracts, architectural decisions | Frontier (Opus) |
| researcher | Codebase exploration, pattern discovery | Fast (Haiku) |
| prompt-writer | Generate bounded, context-rich prompts for sub-agents | Fast (Haiku) |
| supervisor | Track management, sub-agent coordination | Standard (Sonnet) |
| backend-impl | Backend implementation | Standard (Sonnet) |
| frontend-impl | Frontend implementation | Standard (Sonnet) |
| infra-impl | Infrastructure and deployment | Standard (Sonnet) |
| docker-deploy | Container builds, image management, deploy | Standard (Sonnet) |
| db-specialist | Schema, migrations, ORM operations | Standard (Sonnet) |
| cqrs-specialist | Event-driven / CQRS patterns | Standard (Sonnet) |
| tester | Unit and integration tests | Standard (Sonnet) |
| e2e-tester | End-to-end browser tests | Standard (Sonnet) |
| reviewer | Code review and quality | Standard (Sonnet) |
| debugger | Root-cause diagnosis and fix | Standard (Sonnet) |
| security | Security audit | Frontier (Opus) |
| code-quality-auditor | Deep quality analysis | Frontier (Opus) |
| evaluator | Rubric-based critic for the evaluator-optimizer loop | Frontier (Opus) |
| watchdog | Parallel safety monitor, drift detection | Standard (Sonnet) |
| work-recorder | Session documentation | Fast (Haiku) |
File Ownership
<!-- Prevents multi-agent file conflicts during parallel work. [CUSTOMIZE] paths for your layout. -->| Pattern | Owner |
src/domain/**, src/application/**, src/infrastructure/**, src/interface/** | backend-impl |
components/**, app/** | frontend-impl |
*.test.*, *.spec.* | tester |
e2e/** | e2e-tester |
infra/**, k8s/**, Dockerfile* | infra-impl, docker-deploy |
db/**, migrations/**, schema files | db-specialist |
Quality Gates
Blocking (must pass)
- Type check clean (for example
tsc --noEmit). - No secrets in committed files.
- All tests pass.
- No
anyin production code.
Advisory (should pass)
- Coverage at or above target.
- No stray debug logging in production code.
- Dependency audit clean.
Conflict Resolution
When instructions conflict: security > correctness > performance > readability > style.
Custom Instructions
- Claude Code: see
CLAUDE.mdand.claude/. - Cursor: see
.cursorrulesif present. - Other tools: this file is the shared contract.
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.