Agents
Agent-shaped inferlets that loop generation against tools, code execution, or other agents.
Running on the dummy driver?
agent-react and agent-codeact parse the model's output as JSON to extract the next action. The dummy driver returns uniform random tokens, so that JSON parse never succeeds — the loops hit max_steps without firing a tool. agent-swarm's coordination layer (broadcast / subscribe) works correctly; only the message content is random.
| Inferlet | What it shows |
|---|---|
agent-react | ReAct (Reasoning + Acting) loop with tool execution. |
agent-codeact | CodeACT: JavaScript code execution as the action space. |
agent-swarm | Collaborative agent swarm coordinating via messaging::broadcast/subscribe. |
Related guide
- Tool-call parser: hand-rolled and native tool-call loops.
- Inferlet ↔ inferlet messaging: pub/sub and queues — the swarm's coordination layer.
- HTTP and MCP: tool transports for agent calls.