Glossary · Agentic Swarm

Agentic Swarm.

Many agents, one ledger, one gate.

Direct answer

An agentic swarm is a supervised group of coding agents working one job: a partitioner splits it into tasks with verification commands, agents execute in isolated workspaces, liveness is tracked by heartbeat, failures walk a bounded repair ladder, and only verified work reaches an integration branch. The supervisor (not any single agent) owns budgets, placement, retries and the merge gate. Swarms pay off on partitionable work (modules, audits, migrations) and lose to a single agent on tightly coupled edits.

In Anvaya

How We Implement It.

01Agentic SwarmAnvaya Swarm runs process-per-agent under a supervisor: 24 concurrent by default, 256 hard cap, shared/worktree/clone placement, Tier 0 objective verification then a critic, and integration only on anv/integration/<run>, never pushed. A 32-agent soak measured the supervisor at 7.55 MB max RSS.

Questions

Asked About Agentic Swarm.

Q

How is a swarm different from subagents?

Subagents are short-lived helpers inside one agent's turn; a swarm is a supervised fleet of full agent processes with its own state, budgets and integration gate. Subagents isolate context; swarms isolate work.

Q

Does a swarm push code?

No. Approved work lands on an integration branch; pushing is never performed under any policy, including --yolo. A human reviews and merges.

Q

How many agents is too many?

When per-task latency stops paying for itself and verification becomes the bottleneck. Start at the default 24, watch per-agent RSS and tokens, and note that the 256 cap is a supervisor limit, not a hardware promise.

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.