SwarmAI
Human directs. AI delivers.
English | 中文
SwarmAI is a self-evolving Agent OS — every interaction upgrades the system's cognition, not just its templates.
Your AI team, one human directing.
Quick Start
git clone https://github.com/xg-gh-25/SwarmAI.git && cd SwarmAI
cd backend && uv sync && cp .env.example .env # edit with your API key
cd ../desktop && npm install && npm run tauri:dev
macOS (Apple Silicon): Or download .dmg from Releases
Requires: Node.js 18+, Python 3.11+, Rust, uv, Claude Code CLI
📖 Full setup guide: QUICK_START.md
Architecture
┌─────────────────────────────────────────────────────────────┐
│ DELIVERY ENGINES Pipeline · Pollinate · Eval │
├─────────────────────────────────────────────────────────────┤
│ KNOWLEDGE LAYER DDD · Memory · Evolution │
├─────────────────────────────────────────────────────────────┤
│ AGENT HARNESS Context · Sessions · Hooks · Jobs │
└─────────────────────────────────────────────────────────────┘
Core Engines
12 interconnected engines. Each independently valuable; together they compound.
| # | Engine | What It Does | Deep Dive |
|---|---|---|---|
| 1 | Context Management | 11-file prompt architecture, 100K budget, 3-tier ownership | docs |
| 2 | Memory Pipeline | 4-tier persistence: DailyActivity → distillation → compound recall | docs |
| 3 | DDD Cultivation | Self-growing domain knowledge, 7-type ontology, Darwinian decay | docs |
| 4 | Autonomous Pipeline | One requirement → push-ready code. Dual-mode: Full + Goal Loop | docs |
| 5 | Pollinate Engine | One message → multi-format brand content | docs |
| 6 | Self-Evolution | Cognitive L0→L3 patching. 37 corrections, zero class repetition | docs |
| 7 | Self-Healing | Invisible recovery: 5 sensors, auto-respawn, user sees nothing | — |
| 8 | Multi-Tab + MessageStore | Concurrent sessions, phase-gated single-writer, cross-tab isolation | — |
| 9 | Hook System | 21 hooks (17 runtime + 4 lifecycle). Sessions never cold-start | — |
| 10 | Job System | Background intelligence: 13 signal feeds, cron, budget-gated | — |
| 11 | 4-Platform Backend | macOS daemon · Hive (EC2) · Windows · Linux. Compile-time isolation | — |
| 12 | Skills + Channels | 86 skills (lazy/always), Slack gateway, 3-tier permission | — |
The compound loop: Memory → Pipeline judgment → DDD → Evolution → Gates → Memory. Remove one, the rest weaken.
📊 More diagrams: Flywheel · Context · Memory · DDD · Sessions · Jobs · Evolution
Thesis & Design Philosophy
Can one builder + AI operate at team scale? We're testing it live.
- One-shot qualified delivery is the real token optimization. Cheap models iterate 5×, cost more than one correct delivery. Code/content as black box: input → qualified output.
- Division of labor is a compromise for limited human cognitive bandwidth — not an optimal design. One agent, many roles, one knowledge layer. (Sub-agents for adversarial verification ≠ division of labor.)
- Knowledge must eliminate itself. Darwinian decay: 90d unreferenced = retirement. A system that can forget > one that can only remember.
- Evolution is cognitive patching, not data accumulation. We change rules you can
git diff. "Thinks differently" ≠ "knows more." - Quality converges, not just improves. Error classes monotonically decrease. Carefulness doesn't scale. Gates do.
- Sessions are discontinuous. Intelligence shouldn't be. 21 hooks fire between sessions. Gets better through use, not updates.
- If you can't measure it, you didn't build it. OS Eval + golden set + change-triggered. Proves convergence in git.
The compound loop itself is the product. You can't extract one piece and get the same effect.
📖 Full thesis + CLASS A case study + convergence evidence: docs/THESIS.md
📖 Discussion #39: Design Philosophy — Six Pillars
Codebase (~170K LOC)
| Layer | LOC | Entry Points |
|---|---|---|
| Core (spine) | ~10K | session_unit.py, prompt_builder.py |
| Core (extensions) | ~41K | core/ — DDD, evolution, proactive, code intel |
| Skills | ~50K | backend/skills/s_*/ (86 modules) |
| Frontend | ~68K | desktop/src/ — React 19, Tailwind, TanStack Query |
| Tests | ~76K | pytest + Vitest |
Stack: Tauri 2.0 (Rust) · React 19 · FastAPI · Claude Agent SDK + Bedrock · SQLite (WAL + FTS5)
Resources
| What | Link |
|---|---|
| Discussions (45+) | Reading Matrix — Builder 45min · Architect 60min · Leader 30min |
| AI Agent Pitfall Guide | EN PDF · 中文 PDF |
| Design Docs | Platform · Pipeline · Memory · Evolution · Pollinate |
| Contributing | CONTRIBUTING.md |
Contributors
SwarmAI — Human directs. AI delivers.