One requirement in, real committed code out. I give the system a high-level requirement and either attach an existing repository or have a new one created. A team of role-specialised LLM agents, a CTO, a Product agent and three Engineering agents, scopes it, breaks it into tickets with acceptance criteria, implements them inside an isolated sandbox, tests them, reviews them, and merges the approved work onto a real git branch. I act as the CEO: I can watch the whole run live and intervene at the four transitions where a human would actually have a say. Out the other end: a real repository with real commits, and a complete event log of who decided what and why.

Working title WJ's Multi-Agent Simulation of Human Society (the project name on the tracker; confirm before the report)
Supervisor Ziyang Wang

Sections

Stack at a glance

Python orchestration with Pydantic models · five agents over three tiers, one Claude model per tier (Opus 5 for the CTO, Sonnet 5 for Product, Haiku 4.5 for Engineering), set in config/roles.yaml rather than in code · one append-only SQLite event log that doubles as message bus, audit trail and ticket board · FastAPI and WebSockets pushing that log to a live dashboard · one Docker container per ticket attempt, non-root and network-isolated · a real git repo with one branch per ticket, where review approval is the merge. Full reasoning in Architecture Decisions.

Resolved: the visual layer was never an either/or. The 2D dashboard shipped first and is the primary surface; a separate top-down 3D office view is being built alongside it in Three.js, built through Layer 6 of 7, with Layer 7 (live event stream) deferred by design. See ADR-004.

Current state of master branch: the orchestrator, sandbox, evaluation harness and first-pass dashboard are all merged. The dashboard visual redesign (worktree-dashboard-redesign) and the 3D office prototype (worktree-office-sim-prototype) are pushed but not yet merged. 186 tests.


Board: FYP on Trello · Code: WJ-FYP on GitHub · Live: Dashboard · 3D Office Visualization

Project Plan

Task Log

Demo and Assets