Advanced

Agent Swarms

Build sequential agent chains, check entitlement and funded credits, and inspect actual execution results and limits.

Updated · Guide version 1.1

Read as Markdown

The supported swarm executes an ordered chain of model completions. Each agent's output becomes the next agent's input. It does not implement arbitrary parallel map-reduce, debate topologies or external tool permissions from a generic manifest.

Create and run a chain

  1. •Open Agent Swarms.
  2. •Create a named swarm, optionally with a description.
  3. •Select agents and add nodes in the intended order.
  4. •Enter a task input and run the chain only when you have authorized actual credit usage.
  5. •Inspect each returned step and the final output.

Creating a swarm uses POST /swarms with:

{"name":"Research then summarize","description":"An ordered prompt chain"}

Nodes are added separately through POST /swarms/{id}/nodes, with an agent_id. Execution uses POST /swarms/{id}/execute and optional input. A JSON topology with an embedded agents array is not the current creation contract.

Entitlement and limits

The backend checks whether each agent is owned by the caller, approved and free, or covered by a completed unrefunded purchase. A swarm is account-owned. The reviewed limits are 20 swarms per user, 12 nodes per swarm and 8,000 input characters.

Execution requires a configured model provider and enough paid marketplace credits for the chain. The reviewed implementation charges per step and refunds a failed step where its recovery path completes. Do not interpret a saved swarm or an empty result as a successful run.

Troubleshooting and limitations

An empty chain cannot run. A forbidden agent needs an eligible ownership/purchase record. An unavailable response can reflect provider configuration. Check actual execution status and wallet entries after a failure; a process interruption can leave an incomplete run.

These agents perform model completions based on their description; this is not execution of arbitrary downloaded workflow files or a managed secrets/tool runtime. There is no generic max_steps or budget_credits field in the shown creation body.

Next

See Automation, Uploading Agents and Wallet & Credits.