Execution Loop — References

The interleaving of reasoning and action that most agent loops implement, and the feedback structures built on top of it.

The loop structure

• ReAct: Synergizing Reasoning and Acting in Language Models (Yao et al., 2022), arXiv:2210.03629 — the thought, action, observation cycle that nearly every production agent loop is a variation of, including the role the observation text plays in the next step.
• Reflexion: Language Agents with Verbal Reinforcement Learning (Shinn et al., 2023), arXiv:2303.11366 — carrying a written critique of a failed attempt into the next iteration, and the question of what belongs in the next context beyond raw observations.

Instrumenting iterations

• OpenTelemetry — spans, attributes, and context propagation; the vendor-neutral basis for making each iteration individually measurable, including the semantic-convention work covering model and agent operations.

Durable iteration

• Temporal documentation — the determinism and replay model that makes a long-running loop survivable, and the clearest statement of what a loop must give up to be resumable.

No verified reference is recorded for context-truncation policies in agent loops. The strategies in production use — sliding windows, running summaries, projection from a step log — are engineering practice without a published comparative evaluation.