Voss lets AI agents work like an engineering team while keeping every action scoped, budgeted, reviewed, and replayable.
The six primitives
Every run is built from the same six primitives.One run, end to end
The MVP flow
A team is declared once in.voss, then a single goal runs through the cage.
voss team run drives the board with an LLM EM, role workers, and the real Reviewer A and B. Cards run one at a time in the project directory. The run ends with your approve or reject, unless you pass --no-signoff. A swarm is the other multi-agent path.- The EM cannot invent roles outside the declared roster.
- Workers cannot write outside their assigned scope.
- Budget cannot be oversold — it is enforced down the session tree.
- Agents cannot mark their own work Done.
- Done requires independent reviewer evidence.
Swarms
A swarm is the live multi-agent path. A client creates one through thevoss serve swarm routes (POST /swarm, or create_swarm in the Rust SDK), and the default roster is a coordinator, two builders, and a reviewer. The coordinator splits the goal into tasks, each with owned files; two active tasks may not own the same file unless one depends on the other. Builders can be native Voss agents or headless Claude Code and Codex sessions, each working in its own worktree, and writes outside a task’s owned files are denied.
When a task finishes, its work becomes a candidate for review. Candidate-ready means review is pending, not that the work is merged or accepted. Projects with judgments.review: shadow also get an advisory Jev requirement assessment for each candidate.
Every swarm run is recorded as a task graph in the task journal: declared dependencies, scope, criteria, and each worker attempt with its role, model, worktree, and review state. Clients replay it by cursor, and an interrupted run is marked interrupted rather than relaunched.
Team config
Declare roles, scope, budget, and tools in
.voss team{}.Engineering Manager loop
The constrained tech lead that decomposes and integrates.
Board
Columns, WIP limits, and artifact-gated transitions.
Reviewers
The independent Reviewer-A / Reviewer-B split.
Session tree
Per-agent budget fan-out and durable nodes.
Principles
Engineering culture as first-class config.