Skip to main content

Steps — Advanced

Advanced step controls in a conversation-mode Tool are identical to those in a regular Tool. This page exists to call out a few patterns specific to chat-first workflows; for the full reference, see Tool steps — Advanced.

Shared controls​

Open any node in a conversation-mode Tool and switch between the side-panel tabs — they are the same tabs as in a regular Tool:

  • Execution Rules — Trigger rule, Condition, Foreach.
  • Guardrails — safety validators applied to the node's input or output.
  • Limit Override — per-node override of usage and timeout limits.

See Tool steps — Advanced for the full reference.

Patterns for conversation-first workflows​

A few patterns come up specifically in chat / copilot Tools:

  • Routing by intent. Use a Condition on a downstream node to fire only when an LLM classifies the latest user message as a particular intent. This keeps your graph linear and easy to read.
  • Tool use in a conversation. Insert API or KB Search nodes after the LLM classification step, gated by Conditions. Their results flow into a final LLM step that composes the user-facing reply.
  • Foreach over chat history. Rarely useful in conversation-mode Tools — you typically work with the latest user message, not a fan-out over all history. If you do need to process multiple turns, prefer Python code with explicit message indexing.
  • Guardrails on user input. Add a Guardrail node early in the chain to validate the latest user message before reasoning. See Guardrails.
  • Guardrails on agent reply. Add a Guardrail node late in the chain (before Final Output) to validate the assistant reply for PII leakage, toxicity, or competitor mentions.

Reminder for chat agents: In a conversation-mode Tool, the value you map in the Final Output node becomes the chain's reply to the user. Plan your step graph so the variable you want returned is available there.