English | 한국어
This project started from a simple question that emerged after building an earlier ATC prototype: when multiple organizations run their own AI agents and those agents must collaborate on a shared artifact/resource, who gets priority, who decides that rule, and is the decision fair and verifiable?
lex-atc explores an agent orchestration runtime that treats priority and intervention as auditable governance and settlement events — not only as opaque operator discretion.
Background: docs/motivation.md
- Consistency-first: correctness over best-effort progress during contention
- Auditability: every intervention and policy decision should be explainable from logs
- Safety by design: fail-fast config guards and clear mode boundaries
- Not “blockchain for everything”; settlement is a tool, not the product
- Not a generic LLM agent framework; focus is contention control and verifiable coordination
| Mode | Purpose | Backend Required | Notes |
|---|---|---|---|
| Standalone (MSW Simulation) | Demo / simulation | No | MSW + simulated events in the browser. Not 1:1 with production latency/auth/failure |
| Backend Mode | Real runtime | Yes | Uses real API/SSE. Suitable for production-like verification |
flowchart LR
UI[Frontend UI] -->|REST /api| API[Backend API]
UI -->|SSE /api/stream| API
API --> GOV[Governance]
GOV --> ISO[Isolation/Sandbox]
GOV --> SET[Settlement/Dispute]
API --> DB[(DB/Events)]
API --> REDIS[(Redis Pub/Sub - optional/HA)]
| Area | Where | Notes |
|---|---|---|
| UI | packages/frontend |
Monitoring/ops UI, MSW standalone simulation |
| Backend | packages/backend |
API/SSE + runtime (agents/governance/isolation/settlement) |
| Shared | packages/shared |
Shared schemas/types/contracts for API/SSE |
| Full doc | docs/architecture.md | Mode flows and operational request flow |
pnpm install
pnpm dev:standaloneVercel production env:
VITE_ENABLE_MSW=trueVITE_API_URL=/api
Standalone mode requires Service Worker. If SW is blocked, it cannot function.
pnpm install
pnpm dev:backendTo keep wallets stable across restarts:
- set
AGENT_KEY_SEEDandTREASURY_KEY_SEED, or - set
ALLOW_DEV_SEED_FALLBACK=true(development default is true; setALLOW_DEV_SEED_FALLBACK=falseto force ephemeral seeds)
PoW or reputation-based priority can be biased depending on the environment and incentives. We treat utility/entropy-based scheduling as an R&D track and keep a working hypothesis that an entropy-derived signal (capturing uncertainty/diversity/unpredictability) can be a fairer input to scheduling and safety policies in some collaboration settings.
This is a hypothesis to be measured and iterated — not a final claim.
Roadmap and known limitations: docs/roadmap.md
- Docs index: docs/README.md
- Backend deployment: packages/backend/DEPLOYMENT.md
- Frontend deployment/QA: packages/frontend/DEPLOYMENT.md, packages/frontend/QA_CHECKLIST.md
pnpm -w verify
pnpm -C packages/frontend test:e2e