A trajectory, or rollout, records an agent’s task execution across its processing-loop steps. It can include the agent’s plans, tool calls, tool results, and conclusions. Unlike the final result returned to a user, the trajectory shows how the agent reached it, making it useful for debugging, evaluation, and improvement. section 1.3 section 4.5
In an LLMAgent implementation, TaskHandler maintains the rollout as a string. It starts empty; after each sub-step, _format_step_for_rollout() formats that step’s chat history and appends it. Step boundaries make the execution history easier to inspect. The formatter skips system messages, so the rollout is a formatted record rather than an unfiltered chat transcript. section 4.2.3
The rollout also provides context for subsequent steps: get_next_step() uses it when deciding what to do next, and run_step() can include it in the system message. Execution continues until the task completes or a stopping condition, such as a maximum step count, is reached. section 4.2.3 section 1.3
Inspecting a Hailstone rollout
For a task starting at 4, the agent calls next_number on 4 and then 2, stopping when the tool returns 1. The final TaskResult contains 4 → 2 → 1; printing handler.rollout exposes the marked sub-steps and tool interactions used to produce it. section 4.3
=== Task Step Start ===
💬 assistant: My current instruction is 'CALL `next_number` on the current number x=2'
💬 assistant: I need to make the following tool call(s):
{
"id_": "7d529c53-2da2-49c7-a538-abfa8217cd62",
"tool_name": "next_number",
"arguments": {
"x": 2
}
}.
🔧 tool: {
"tool_call_id": "7d529c53-2da2-49c7-a538-abfa8217cd62",
"content": "1",
"error": false
}
💬 assistant: The `next_number` function returned the number 1 when called with x=2.
Now, I will provide the complete sequence observed from start to finish: Starting at x = 4 and ending with x = 1.
Here is the sequence:
```
4 -> 2 -> 1
```
Since we received the final number 1 as requested by the instruction, no further calls are needed. The task can now be considered complete.
=== Task Step End ===Inspect the second sub-step’s instruction, tool-call arguments, result, and completion message. section 4.3
Long rollouts consume context-window space. A subagent can execute a subtask with its own isolated rollout and return only the result to the delegating agent. Rollouts can also expose interactions absent from the final answer, such as a remote agent requesting additional input and a subsequent call to resume its task. section 9.1.1 section 10.2.3
























