Agents work through a goal on their own, using the tools and workspace you give them. Each one is a system prompt, a model policy, a toolset and a bounded loop that keeps going until the task is done or a limit is reached.
A bounded plan-act-observe loop
Every run has a step cap, a token budget and an abort signal, and every stopping reason is reported rather than hidden. That is what makes an autonomous loop safe to leave running: it cannot spend without limit, and it always says why it stopped.
Sub-agents and swarm mode
An agent can delegate to other agents through a depth-capped call tool with a shared scratchpad, which is how a researcher to writer to critic chain works. In swarm mode, work fans out to several member agents in parallel and their results are synthesised, with a per-member trace you can drill into.
A full step trace
Every run shows its complete step-by-step trace, not just the final answer, including each tool call and each observation. An autonomous loop that only shows its conclusion cannot be trusted or debugged, so the trace is always available.