Agent Episode Prompt Stack

Compose an agent episode as fixed system, optional demonstration, instance brief, then per-turn next-step templates around action and observation.

An interactive agent episode is easier to reason about as a prompt stack than as one growing blob. SWE-agent's default order:

  1. System — fixed for every instance; never collapsed. Describes the environment, documents custom commands (usage + docstring each), and locks the response format (typically one thought plus one action per turn).
  2. Demonstration — optional successful trajectory. Either inject each demo turn as separate history messages or wrap the whole demo in an explicit “this is a demonstration” frame; which representation helps is model-dependent. Demos mainly teach command format more than domain strategy when tasks are diverse.
  3. Instance — the concrete task plus brief reinstruction, often with failure-derived prompt tips.
  4. Next step — after each well-formed action, append observation (or an explicit silent-success message when stdout is empty) plus the shell/prompt cue for the next command.

The turn loop is LM inference → execute → next-step template, until submit, budget exhaustion, or action format enforcement stops the episode. Keep system content instance-agnostic; put task text only in the instance template so the stack stays reusable under system prompt architecture and conversational agent context.

Sources

  • SWE-agent (Yang et al.), pp. 31–45