What an agent swarm is
One agent works through a task one step at a time. A swarm splits the task into pieces, gives each piece to its own agent with a fresh context, and merges what comes back. For work that is wide, such as researching 200 companies or checking 80 websites, that turns hours of sequential work into minutes, and each agent stays focused on one thing.
The hard parts are not the model calls. They are giving every agent a safe place to run, a browser that works on the real web, a way to receive email, keeping the run inside a budget, and checking the results before they reach you. That is what Swarm Computers runs for you.
How a run works
- 01
Describe the task
Plain language, a size (Small, Medium, Large or Max), a budget cap and an optional deadline.
- 02
Plan and estimate
The coordinator returns a plan and a cost range before any agent starts. Approve it, or let it start on its own when it fits the cap.
- 03
Rounds of work
Agents fan out, then review each other's results and fill gaps. Spend and progress stream live.
- 04
One result
The merged result and every file the agents made land in the run's shared folder.
What every agent gets
- Its own virtual machine, never shared with another customer.
- A real browser for sites that need logins, clicks and forms.
- Its own email inbox, for sign-ups and confirmations.
- Access to the run's shared folder, so agents can build on each other's work.
- An uncensored model, so legitimate work does not stall on a refusal.
Call it
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
}'With plan_only you get the plan and estimate back first and start the run when you are happy. The request shape is in preview and may change before general availability.
Good fits, and poor ones
| Good fit | Poor fit |
|---|---|
| Research across many companies, products or sources | One tightly coupled task, such as a single tricky bug |
| Checking many websites or pages the same way | Work with no way to check the answer |
| Testing your own product as many different users | Anything aimed at systems you do not own or have permission to act on |
| Security testing of domains you have verified |
Questions
- How many agents will run?
- You choose a size, which sets how many agents may work at once: up to 5, 20, 50, or 100 for Max. The coordinator decides how many the task needs within that.
- How is it billed?
- One all-inclusive rate per agent-minute, covering model, VM, browser, inbox and storage. You are billed for minutes actually used and never above your cap. See the pricing page.
- What happens if the budget runs out?
- You get a warning at 80%. At the cap the run pauses with a partial result, and you can add budget within 7 days to finish it.
- Who can use it?
- Agents with browsers and inboxes are powerful, so accounts are reviewed before their first run.
Related
- Agent swarms, explained and on tapWhat an AI agent swarm is, when a swarm beats a single agent, and how to run one through an API: many agents in parallel, each with its own computer, one merged result.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
- A sandbox for every agent, includedEvery agent in a Swarm Computers swarm runs in its own sandbox: a separate virtual machine with a real browser and its own inbox, deleted when the run ends.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