Skip to main content

​
Introduction

On a Guru that has agentic capability enabled, each question can be answered in one of two modes, chosen with the Chat | Agent switch above the question box: Both modes can use your Actions and MCP tools — the difference is how.
  • Chat (the default): a single grounded pass. Your Guru retrieves the relevant sources and, if a tool clearly matches the question, runs it once before answering. Fast, predictable, and inexpensive.
  • Agent: an autonomous multi-step loop. The model decides which tools to call, can chain several across rounds, and reasons between them before answering. More capable for multi-step tasks, but slower and more expensive.

​
Chat vs Agent

​
When to use which

  • Reach for Chat for the large majority of questions: definitions, lookups, “what does the docs say about X”, or a single action (“what’s the current status of Y”). It is the fast, cheap default and it still uses a directly-relevant tool.
  • Switch to Agent when the question is really a small task: “find the failing simulation, pull its latest run, and summarize the top threats”, or anything that needs the Guru to gather data across several tools and reason over the results.
Agent mode is a per-question choice. Selecting it applies only to that question (and its follow-ups in the same chat); starting a new chat returns to Chat.

​
Availability

The Chat | Agent switch appears only when a Guru is agent-capable: multi-step answering is enabled for the Guru and it has at least one usable tool (an enabled Action or a connected MCP server that exposes tools). Gurus without tools, or with multi-step answering off, always answer in Chat and show no switch. Chat is always the default. Even on an agent-capable Guru, ordinary questions stay fast and cheap unless you explicitly pick Agent.

​
Using it from the API

The api/v1 answer endpoint accepts an optional mode field:
  • "chat" (default) — single-pass answer.
  • "agent" — the autonomous multi-step loop.
agent takes effect only on an agent-capable Guru (multi-step enabled + at least one tool); on any other Guru the request falls back to Chat.
Omit mode (or send "chat") for a normal fast answer.
Agent answers are slower and cost more than Chat answers, because the tool loop runs before the first answer token and each round re-sends the conversation plus the tool schemas. Use Agent when the task needs it; leave ordinary questions in Chat.