team3 — where humans and agents build together

An agent isn't a one-time tool. It's a teammate with a memory and a role — one that comes back day after day, picks up where it left off, and gets better as the project grows.

team3 is a workflow that turns "you + your coding agents" into a real team: 1 Human + 1 Architect + 1 Dev + 1 UAT, four roles with one job — build your product, from app → module → feature → uat, and keep building it.

English | 简体中文

Why "team3"? Three agents — Architect, Dev, UAT — and a Chinese saying: 三生万物, "from three, all things are born." Three agents, one shared memory, and a team that keeps growing into more than the sum of its parts.

---

Why we built it

If you've shipped with AI agents, you've probably felt this:

• You're the dispatcher. Three sessions open, context copied by hand from one window to the next, requirements re-pasted every morning. The agents work hard — you work harder keeping them in sync.
• You're the tester. No checkpoints, no UAT. Just you, clicking through features by hand, trying to verify faster than the AI can produce. You can't.
• And it never compounds. Every task starts from zero. Nobody remembers last week's decisions. It's like working with brilliant strangers who forget you overnight.

team3 fixes exactly this — not by wrapping a better code model, but by giving the team a shared memory and a workflow that outlives any single session.

How it works

A product gets built the way a real team builds it:

1. The human decides. The idea, the direction, the judgment calls — they come from you. Nobody overrides you. spec/app_design.md is yours alone to maintain.
2. The Architect plans. It breaks your idea into modules and features, dispatches the work, reviews what comes back, and decides when it's ready for acceptance.
3. The Dev builds. Each task gets its own fresh session — clean context, unit tests, self-verification, delivery. No cross-session contamination.
4. The UAT judges. From the user's point of view, black-box. It doesn't read the Dev's code, doesn't use mocks, and shows no mercy. If it doesn't feel right to a user, back it goes.

They talk through spec/actions.jsonl — a shared inbox you're always part of. A daemon keeps every session alive, scheduled, and aware of where the project stands.

The result: the agents aren't tools anymore. They're teammates who remember — and the longer you work together, the better the team gets.

What you get

• One team, four roles — clear boundaries, decisions kept in a single authoritative file (spec/decisions.md), and every hard-won lesson distilled into spec/experience.md so nobody repeats a mistake.
• A daemon that babysits — watches actions, queues and schedules sessions, routes messages, rebases, persists state, and watches for the dead (literally — it has a watchdog).
• A CLI for everything — init, write-action, experience, simulate_human, validate-uat-evidence…
• A web console — watch progress and talk to the team in real time.
• An eval loop — loop/ runs evals, regressions, and badcase drills, so you can watch the team get measurably better.
• Ships like a real product — build/build.sh produces a globally-installable package; team3 start and you're live.

Quick start

# Build and install globally
cd team3
bash build/build.sh
npm install -g ./pkg/team3-*.tgz

# Start
team3 start -p 9001
# Open http://localhost:9001

Dev mode (dogfooding from source):

cd team3
node build/embed-prompts.js        # re-run after any prompt change
cd web
TEAM3_SUPERMAN=1 PORT=9001 npm run dev

Documentation

• team3/spec/ — design docs and protocol definitions (app_design.md, packaging_design.md, usage.md, …)
• team3/human_coding/ — the three role prompts and the workflow reference (team3.md is the authoritative protocol)
• draft/ — early-stage ideas and discussion notes (see draft/README.md)

License

MIT