Archon Protocol

English

中文

English

中文

Appearance

Archon ProtocolAI Architect Evolution Protocol

The operating system for AI agents — kernel, drivers, syscalls, filesystem. One protocol governs all.

Get Started
Design Philosophy
Architecture
🧠

Kernel: Always Resident

AGENTS.md is the kernel image — always in context, never paged out. It defines identity, workflow, and constraint enforcement. Everything obeys it.

🔌

Drivers: Constraint Skills

Hard boundaries loaded into kernel space via `skills:` field. ❌ prohibitions are law, not suggestions. Grep-verifiable, CI-testable.

⚡

Syscalls: 4 Commands

boot() init → exec() demand → stat() audit → defrag() refactor. Each triggers a well-defined kernel operation.

🔧

Hardware Detection

Auto-detects Cursor, Claude Code, Codex, Copilot, Windsurf, Gemini CLI. Deploys drivers to the right paths.

🧬

Directed Evolution

Every task strengthens the constraint system. Anti-patterns become prohibitions. Quality monotonically increases.

📂

Filesystem: Persistent Memory

Architecture = /usr/src/, guides = man pages, ADRs = /var/log/, proposed-rules = staging repo. Clear mount semantics.

What Is AAEP? ​

AAEP (AI Architect Evolution Protocol) defines how an AI agent evolves into a project's architect through continuous constraint → execution → evolution cycles.

AI coding tools provide capability — code generation, file operations, terminal execution. They don't provide methodology — workflow governance, quality assurance, experience accumulation.

AAEP is the methodology layer.

The Four Layers ​

LayerWhatHow
ConstraintHard boundaries that shape code generation❌ prohibitions in constraint skills, injected into agent context
WorkflowStandard delivery sequences/archon-demand 7-stage pipeline with opt-out flags
EvolutionSystem that gets smarter with every taskStage 3.6 discovers patterns → staging area → approved → new constraint
KnowledgeProject memory that persists across sessionsarchon.config.yaml, constraint skills, architecture docs

The Feedback Loop ​

Task 1 → Agent writes code under constraints
       → 6-dimension self-audit catches issue X
       → Proposes: ❌ prohibit X
       → User approves → constraint added

Task 2 → Agent cannot make mistake X (it's in the constraint set)
       → Audit catches issue Y → same cycle

Task N → Constraint system is comprehensive
       → Code quality monotonically increases
       → Fewer fixes → faster delivery

More tasks → better constraints → higher quality. This is directed evolution — always toward fewer bugs, better performance, more consistent architecture.

Why Not Just Use ESLint? ​

DimensionESLintAAEP
No any type✅✅
File size limits⚠️✅
Async: skeleton + error + retry per section❌✅
Tests updated for changed signatures❌✅
Module boundary violations❌✅
Off-screen deferred with IntersectionObserver❌✅

ESLint catches syntax. AAEP catches architecture. Both layers active — Stage 1.5 runs your linter automatically.

Cross-Tool Compatibility ​

ToolAgentsSkillsConstraints
Cursor✅ (primary)✅✅ (preloaded)
Claude Code✅ (primary)✅✅ (preloaded)
Codex—✅✅
Copilot—✅✅
Windsurf—✅✅
Gemini CLI—✅✅
20+ others—✅✅

One protocol, every tool. Agent-first where supported, skill-fallback everywhere else.

Powered by AAEP (AI Architect Evolution Protocol)

Archon Protocol