Skip to content

Agent Reasoning

Agent Reasoning is a guided explanation panel for reviewers. It is not meant to expose every internal token or raw implementation detail. It should make visible agent behavior understandable.

Two trace types

LLM decision trace

An LLM decision trace appears when the backend asks the model to choose an action.

It can show:

  • resident,
  • location,
  • time,
  • nearby NPCs,
  • memory summary,
  • prompt summary,
  • raw model response,
  • parsed action,
  • reason.

Speech-sync trace

A speech-sync trace explains the same NPC_SPOKE event that drives a visible map bubble.

It is not a separate LLM call. It helps answer:

  • Which visible speech bubble is this?
  • Who spoke?
  • What was said?
  • Why did the system display it?
  • Was it from autonomous behavior, visitor chat, local ambient speech, or a visitor task?

What reviewers should learn

Agent Reasoning should show that the system is designed as a multi-agent environment:

  • each resident has identity and goals,
  • each resident has memory,
  • residents perceive nearby context,
  • visitor requests can become tasks,
  • actions are broadcast as real-time events,
  • visible map behavior is linked to traceable system events.

What not to overclaim

Do not present every speech-sync row as a separate LLM decision. Some lines are local ambient speech or synchronized display traces. The panel should distinguish these clearly.