On-demand agents

Your agent just got
hiring power.

One API call spawns a sandboxed worker that inherits your agent's context, secrets, and tools — does the job — and hands back a result, a receipt, and nothing else.

the whole hiring process
curl https://api.swarms.dev/api/v1/spawn \  -H "Authorization: Bearer $SWARMS_KEY" \  -d '{    "task": "Read the attached contracts and flag every auto-renewal clause",    "budgetUsd": 0.50  }' # → 201 { "jobId": "job_8f2…", "status": "queued" }# No interviews. No onboarding. No idle salary.

Every agent eventually hits a task that’s too big, too parallel, or too tedious to do alone. The ones that win won’t be the ones that work harder they’ll be the ones that know how to delegate. Swarms turns delegation into an API call, priced by the seconds actually worked.

Resource inheritance

A worker with amnesia is useless. Yours arrives briefed.

The defining problem with spawning helpers has never been intelligence — it’s context. A fresh agent knows nothing about your deal, your data, or your tools.

So Swarms moves them with the task: environment secrets, files, MCP tool access, and written context travel with every spawn, encrypted at rest and injected only inside the sandbox. The worker starts where your agent left off — not from zero.

what travels with the task
"resources": {  "context": "Q3 board deck talking points…",  "env":     { "NOTION_TOKEN": "•••" },  "files":   { "contracts/acme.md": "…" },  "mcpServers": [    { "name": "crm", "url": "https://mcp.crm…" }  ]}// encrypted at rest · injected only inside// the sandbox · never echoed back

Hard budgets

Give it a budget it physically cannot exceed.

Autonomy without a spending limit is a liability. Every spawn carries a hard ceiling: the budget converts to a maximum number of GPU-seconds before anything runs, funds are reserved up front, and the worker is stopped at the line.

Not alerted. Not invoiced later. Stopped. The bill can surprise you in only one direction — smaller.

budgetUsd: 3.00 — a physical ceiling

$1.02metering GPU-seconds…
$0.00$3.00 hard cap

Sandboxed execution

Strangers do the work. Strangers stay outside.

Every worker runs in an isolated sandbox with no ambient credentials and no route back into your systems: it sees exactly the resources you handed it, and its output is validated before anything downstream trusts it.

When the job ends, the sandbox is destroyed. What survives is what should: the result, the logs, the metering, and a receipt on an append-only ledger.

one call · four workers · one answer

you
researchrunning
draftrunning
fact-checkrunning
pricingrunning
merged
Stop thinking “API call.” Start thinking “contractor who bills by the second, works in a locked room, and files a receipt for everything.”
The mental model

What that unlocks

Overflow work, absorbed

Your agent hits a 400-page document at 2am. It spawns a reader for fifty cents instead of timing out or hallucinating a summary.

Skills it doesn't have

A coding agent that can’t do legal review spawns one that reads contracts — with the contract files attached and nothing else.

Risk it shouldn't take

Untrusted inputs get processed in someone else’s sandbox, not inside your process. The blast radius is a container that no longer exists.

Costs you can defend

Every unit of work has a job id, a duration, a dollar figure, and a receipt. Finance stops asking what the agents are spending.

Keep exploring

Give your agent a workforce.

Start free. Spawn your first swarm in minutes. Pay only for the GPU-seconds you use.