Gym icon indicating copy to clipboard operation
Gym copied to clipboard

Docs: Multi-Step Patterns

Open cwing-nvidia opened this issue 1 month ago • 1 comments

Document Multi-Step Patterns

Context

Users have asked how to handle multi-step agentic tasks where the agent makes multiple tool calls within a single trajectory. This requires understanding:

  • How NeMo Gym maintains state across multiple steps
  • The session model and how it routes requests
  • When to use stateful vs. stateless patterns
  • How to manage state persistence (in-memory vs. database)

We should add documentation for this foundational pattern, and update the README for the example_stateful_counter resource server which demonstrates this.

Problem

Multi-step interactions are core to agentic tasks, but users don't have clear documentation on:

  • What happens when an agent makes multiple tool calls
  • How to maintain state between those calls (e.g., keeping track of what the agent has done so far)
  • The session lifecycle and routing model
  • Best practices for stateful resource servers

Acceptance Criteria

  • [ ] Create a "Multi-Step Interactions" section in documentation that covers:
    • What multi-step means in the context of NeMo Gym
    • How sessions work (creation, routing, cleanup)
    • When you need stateful vs. stateless resource servers
    • State management patterns (in-memory, database, external services)
  • [ ] Document example_stateful_counter as a reference implementation:
    • Write comprehensive README
    • Code walkthrough of key patterns
    • Setup and testing instructions
  • [ ] Add multi-step pattern to main tutorial flow or as prominent "next step"
  • [ ] Consider adding a decision tree: "Does my task need state management?"

Priority

High - foundational pattern needed for most agentic tasks

Related

  • Database integration tutorial will build on these session patterns. We need to determine how we structure database examples: as part of this multi-step documentation or as standalone documentation (or both, given this is commonly requested).
  • This is separate from multi-turn conversations (user <-> assistant exchanges)

cwing-nvidia avatar Nov 12 '25 19:11 cwing-nvidia

@lbliii tagging for visibility

cwing-nvidia avatar Nov 12 '25 19:11 cwing-nvidia