Definition
An agentic swarm is a set of specialized LLM agents (and tools) working on the same goal, usually in parallel, with explicit coordination, verification gates, and a shared paper trail.
If “multi-agent” is many agents exist, a swarm is many agents behave like a system.
What it looks like in practice
A swarm is usually not “10 agents chatting.” It’s role separation and handoffs:
- An explore agent maps the territory, finds risks, and proposes a bounded plan.
- One or more task agents implement specific slices.
- A “critic” pass (often oppositional validation) tries to break the work.
- Everything emits a build receipt so humans (and other agents) can verify, not trust.
Why it matters
Swarming is a throughput play and a correctness play:
- Parallelism gets you speed, but role separation gets you fewer self-inflicted wounds.
- Verification gates reduce the “looks right” bias of a single model run.
- Receipts make outcomes auditable (and mergeable) without babysitting.
What it’s not
- A consensus trick (“have 5 agents vote”).
- A substitute for boundaries (permissions, sandboxing, secrets).
- An excuse to skip gating (“it ran in my head”).
Minimal swarm checklist
- Clear role boundaries (who explores, who edits, who verifies).
- A shared source of truth (issues, checklists, acceptance criteria).
- A gated multi-agent flow so progress is earned, not narrated.
- A place to run safely (shadow fork or sandbox environment).
- A coordination loop (AgOps): costs, rate limits, retries, and logs.
Practical rule
If you can’t explain who verifies each agent’s output, you don’t have a swarm—you have a hope.