Docs · Agent Swarms (Orchestration & Review)

Agent Swarms (Orchestration & Review).

Partition, isolate, verify, integrate.

Direct answer

A swarm runs one job across many agent processes under a supervisor: a partitioner splits it into tasks with mandatory verification commands, agents execute in shared, worktree or cloned workspaces, and only verified work reaches an integration branch. Start with a small partition on a repo with a clean git tree; the supervisor handles liveness, retries, budgets and the merge gate.

Agent Swarms (Orchestration & Review)

Steps

011 · Validate a manifestanv swarm validate --file run.json checks the plan offline (tasks, placement, verification commands, budgets) before anything spawns.
022 · Start the runanv swarm up --file run.json launches the supervisor detached. Run state and artifacts appear under .anvaya/swarm/ in the repository.
033 · Watch the fleetanv swarm status (or --watch, or --json for CI) shows per-agent state, RSS and tokens; anv swarm logs <task-id> streams a task's output; the TUI Swarm screen is the live board.
044 · Handle gates and failuresanv swarm approve/reject resolve framework gates; retry relaunches a terminal task; hold/release control admission by wave; cancel stops a task gracefully (SIGTERM then SIGKILL on the process group).
055 · Review the integration branchApproved work lands on anv/integration/<run>. Nothing is ever pushed, including under --yolo, inspect the branch, run your own tests, then merge on your terms.

Agent Swarms (Orchestration & Review)

Knobs that matter

01Placementshared is cheapest and fine for disjoint files; worktree gives each agent its own checkout with shared git objects; clone is the strongest isolation. The defect, MVP and refine bundles default to worktree.
02Concurrency24 agents concurrently by default, 20 spawns per second, 256 hard cap. Raise only while per-task latency still pays for itself.
03VerificationEvery partition task carries a verification command that must pass before the task counts as done; failures walk a bounded repair ladder instead of retrying blindly.
04LivenessA 10-second heartbeat drives alive/suspect/lost states; a hung agent is cancelled and its process group cleaned up rather than blocking the run.

Questions

Asked About Agent Swarms (Orchestration & Review).

Q

Is anything pushed automatically?

No. Integration happens only on anv/integration/<run>; pushing is never performed under any policy. You review and merge.

Q

When should I not use a swarm?

Tightly coupled edits in one file, or work without an objective test. Two agents editing the same symbols usually lose to one agent plus a checkpoint.

Q

How do I control cost?

Partition conservatively, keep verification at the cheapest effective tier (linters and tests before model critics), cap repair attempts, and watch the per-agent token column on the board.

Run Agents That Fit On Your Laptop.

25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.

Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.