A procedural agent is an LLM program that takes conversation data and events as input (either a completed conversation when run post-hoc, or in-progress when called mid-conversation). It uses its instructions and this input to execute procedures and output structured objects. Unlike conversational agents optimized for real-time responses, procedural agents can utilize extended reasoning time for complex analysis and multi-step execution. Conversational AgentReal-time interactionFast responsesCustomer-facingProcedural AgentTask-focusedExtended thinkingStructured output

​
Key capabilities

Procedural agents excel at tasks that require:
  • Extended reasoning: Can use sophisticated thinking models that take time to process complex scenarios
  • Multi-step planning: Break down complex tasks into sequential steps and execute them methodically
  • Structured outputs: Generate specific data formats, reports, and actionable recommendations
  • Deep analysis: Process large amounts of context (like full conversation transcripts) to extract insights
  • Tool orchestration: Chain multiple tool calls together to accomplish complex objectives
Procedural agents can utilize extended thinking time (sometimes tens of minutes) for thorough analysis, unlike conversational agents optimized for sub-second responses.

​
Core use cases

​
Post-hoc analysis

Triggered automatically when conversations terminate with specific outcomes. The agent receives the complete conversation transcript and event log as input, processes this data through its instruction set, and outputs structured objects for downstream systems.

​
Tool-based execution (mid-conversation)

Invoked synchronously by conversational agents during active sessions. The procedural agent receives the current conversation state and events up to the invocation point, executes its procedure, and returns structured output to the calling conversational agent. Conversational agent makes tool call @procedural_agent_name, waits for JSON response, then continues conversation flow with the returned data.

​
Standalone execution

Triggered via API calls or scheduled events, independent of conversation context. These agents execute complex workflows and produce detailed structured outputs for business operations.

​
Technical architecture

Procedural agents use the same underlying infrastructure as conversational agents but with key differences optimized for task completion rather than real-time interaction. They can utilize extended thinking models that take time to reason through complex scenarios, and include built-in output recording tools for generating structured data and reports. Configuration options allow you to control planning intensity—how much upfront analysis the agent does before executing tasks—and specify output formatting requirements. Like conversational agents, they use the same tool-calling mechanisms and have access to customer context and memory.

​
When to use procedural agents

Common use cases for post-hoc analysis:
  • Extract actionable next steps for the human ops team
  • Identify high-priority follow-ups and notify the team over Slack
  • Compute quality and success metrics based on the conversation transcript
Common use cases for tool-based execution:
  • Given the conversation so far and a set of policies, compose a CRM ticket
  • Based on the available transfer lines with ops teams, decide where to transfer the customer