liveBook · Manning

Run()

As explained in Build a Multi-Agent System (from Scratch)

run() is an LLMAgent method that accepts a user task, initiates its processing loop, and returns an awaitable that resolves to a TaskResult.

LLMAgent.run() submits a Task to an agent’s processing loop. It creates a TaskHandler and schedules the loop as a background asyncio.Task. Although run() is not itself an async coroutine, it returns the handler, an asyncio.Future subclass that can be awaited for a TaskResult. section 4.1 section 4.5

A Task carries an id_ and instruction; its TaskResult carries a task_id and content. The initial interface accepts an optional max_steps limit, allowing callers to bound the number of sub-steps. section 4.1 section 4.2.3

The processing loop delegates execution to TaskHandler:

  • Call get_next_step() with the previous TaskStepResult, if one exists. It returns either a TaskStep or a completed TaskResult. section 4.2.3
  • For a TaskStep, call run_step() and use its result in the next cycle. Each sub-step’s chat history is formatted and appended to the handler’s rollout. section 4.2.3
  • Finish when a TaskResult is produced, or stop on a condition such as an irrecoverable error or the maximum sub-step count. section 4.2.3

Callers can await agent.run(task) directly, or retain the handler to inspect its rollout and check done(). Calling handler.result() before completion raises InvalidStateError. section 4.3

Run a tool-assisted task

For a Hailstone task starting at 4, submit the Task with handler = llm_agent.run(task, max_steps=5). After the handler completes, handler.result() yields a TaskResult whose content is 4 → 2 → 1. The rollout records the calls to next_number for 4 and 2. section 4.3

Extensions add capabilities at task-run boundaries. Skill discovery occurs when the handler is constructed, with skills_scopes and explicit_only_skills forwarded by run(). With configured memories, the loop loads recalled episodes and records task execution. The optional with_approval flag requests human review of a proposed result; rejection feeds operator feedback back into the loop. section 6.3.3 section 7.5.2 section 8.3.1

Figure 8.9 The sequence of interactions between a human operator, LLMAgent, and the TaskHandler it dispatches for carrying out the Result-Approval HITL pattern

Trace the handoff from task execution to human review and the approval or rejection paths. section 8.3.1

How Build a Multi-Agent System (from Scratch) covers it

How deep it goes
Implements the background processing loop, step selection, execution, stopping conditions, and extensions for skills, memory, and approval.
Examples
Uses Hailstone sequence tasks to demonstrate bounded execution, rollout inspection, and human approval of a proposed result.
What you can do after
Build it from scratch as a background processing loop returning an awaitable TaskHandler.
How it connects
Connects agent processing loops with tool calling, asynchronous futures, skill discovery, episodic memory, and human-in-the-loop approval.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build a Multi-Agent System (from Scratch). The section links point to the parts of the book it draws on.