A team is not a list of prompts
AI Teams treats coordination as a product and systems problem. Work lives in dependency-aware issues; every run has identity, state and cost; heartbeats keep the process alive; and persistent interactions allow work to pause, request context and resume without pretending continuity.
Lead
Interprets the objective, proposes the plan, hires roles and remains accountable for closure.
Specialists
Engineer, Reviewer, QA and other roles receive deliverables, boundaries and an explicit reporting line.
Control plane
SQLite keeps the durable queue, activity, budgets, reviews and recovery state.
Adapters
Codex, Antigravity and other runtimes connect without confusing installation, authentication, health and authority.
Less ceremony when risk is low
The system does not assume a full team is better. It can use solo_lead, lead_quorum or full_team according to risk and evidence. Heavy coordination should disappear for simple work; independent review, hard gates and escalation may become mandatory for critical changes.
- ✓One root issue with acceptance criteria and an exit condition.
- ✓Traceable delegation and roles with bounded authority.
- ✓Deterministic gates an LLM cannot override.
- ✓Budget and liveness limits defined before execution.
- ✓Accepted closure, explicit blocking, or evidence explaining non-convergence.
Evaluation is part of the product
The workflow_v2 benchmark compares a direct agent, Solo Lead, Lead Quorum and Full Team on the same case. It seals the model, adapter, commit, worktree, seeds and hidden suites; separates convergence, quality and economics; and requires blind judgment plus a human sample before recommending policy.
Public status and boundary
The repository is public under Apache-2.0 and control-plane installation is verified on Windows x86_64 through a reproducible clean-room run. Linux and macOS remain unverified targets; the declared version is 0.1.0 and there is no stable release yet. Newer local work remains in development and is not used as public proof here.
- ✓Public control-plane code and documentation.
- ✓Windows bootstrap verified in GitHub Actions.
- ✓Documented Lead-first architecture and SQLite persistence.
- ✓Conservative evaluation with visible failures.
- ✓No claim of full autonomy, complete multi-platform support or guaranteed savings.