Skip to main content

Planner Models and Backends

agx delegates planning to an LLM backend. This document describes how planner models are used and how backends can be swapped.

Responsibilities of the Planner​

Given:

  • an input payload (stdin),
  • a natural-language intent,
  • a tool registry (available tools + --describe metadata),

the planner must produce:

  • a valid JSON plan matching /specs/plan.schema.json,
  • safe, deterministic tool invocations,
  • minimal, composable steps that can run in local or distributed mode.

Backends​

Examples of supported/planned backends:

  • Local LLMs via Ollama.
  • Fine-tuned models via Tinker.
  • Remote APIs (OpenAI, DeepSeek, Anthropic) where appropriate.

All backends must implement a common trait so agx can switch planner models without changing core logic.