What an agent swarm is
A single agent works through a task step by step in one long conversation. As the task grows, its context fills up and it slows down. A swarm takes the same task, has a coordinator break it into independent pieces, and gives each piece to a separate agent with a fresh context. When the agents finish, their reports are checked and merged into one result.
The name comes from swarm intelligence in nature, where many simple workers produce useful group behaviour. In practice most useful agent swarms are less mysterious than that: a planner, many workers, and a merge step.
Common swarm patterns
| Pattern | How it works | Good for |
|---|---|---|
| Fan-out and merge | A planner splits the task; workers run in parallel; a merger combines reports | Research across many sources, checking many sites |
| Rounds | After each round a reviewer decides what is missing and starts more agents | Open-ended tasks where gaps only show up later |
| Shared board | Agents post short notes every other agent can read | Avoiding duplicate work and passing on leads |
| Handoffs | One agent passes a conversation to a specialist agent | Support and triage flows |
Swarm Computers uses the first three: fan-out with a coordinator, rounds with a reviewer, and a shared board.
When a swarm beats one agent
- The work is wide: 200 companies to profile, 80 sites to check, 50 user journeys to try.
- The pieces are independent enough that agents do not need each other's answers to start.
- Each piece needs real tools, such as a browser or an inbox, and takes minutes, not seconds.
- You can check the result, because a swarm is only as good as its merge step.
If the task is one tightly coupled problem, such as a single subtle bug, one strong agent is usually better.
Running a swarm with Swarm Computers
curl -X POST https://swarmcomputers.com/v1/swarms \
-H "Authorization: Bearer $API_KEY" \
-H "Content-Type: application/json" \
-d '{
"task": "Profile these 200 suppliers: pricing, certifications, contacts. One row each, with sources.",
"size": "large",
"budget_usd": 150,
"plan_only": true
}'Choose a size, which sets how many agents may run at once (5, 20, 50 or 100), and a budget cap. The coordinator returns a plan and estimate, runs agents in rounds, and stops new work as the cap gets close. You are billed per agent-minute actually used and never above the cap.
Questions
- What is the difference between an agent swarm and a multi-agent system?
- A multi-agent system is any setup with more than one agent. A swarm is a multi-agent system built for scale: many similar agents working in parallel on pieces of one task.
- How many agents can a swarm run?
- With Swarm Computers, up to 100 at once on the Max size. The coordinator decides how many the task actually needs within the size you choose.
- Do I need a framework to run an agent swarm?
- Not with Swarm Computers. Frameworks help you write the orchestration code, but you still have to host the agents, their browsers and their sandboxes. Swarm Computers runs all of that behind one API call.
- Are agent swarms expensive?
- They use more agent-minutes than one agent, but finish sooner. With Swarm Computers you set a hard budget cap on every run, see an estimate first, and get a warning at 80%.
Related
- A swarm of AI agents behind one API callSend one task and a budget; a swarm of AI agents works it in parallel, each with its own VM, browser and inbox. One merged result, never billed above your cap.Read more
- Multi-agent orchestration, hostedMulti-agent orchestration without the infrastructure: planning, parallel agents in their own sandboxes, budget limits and a merged result, behind one API call.Read more
- Swarm computing, for AI agentsSwarm computing means many simple workers solving one problem together. Here is the classic idea, how it applies to AI agents today, and how to use it through one API.Read more
- Two APIs. Two simple prices.Two prices: the Model API per token, and the Swarm API at one all-inclusive rate per agent-minute with a hard budget cap. Prices provisional.Read more