ACE: structured context refined through performance feedback

Agentic Context Engineering (ACE) incrementally accumulates and refines structured context items based on performance feedback. The source presents this as an alternative to iterative prompt rewriting and compares it with the GEPA approach.

To study or apply the idea, select a complex task and separately record the context items used and the observed results. This makes it easier to assess what helped without attributing an improvement to a particular change without evidence.

Make changes gradually and keep a representative set of test cases. Compare results before and after each change; useful feedback should be tied to clear performance criteria, not just general impressions.

If you use AI to analyze internal examples, remove or replace personal data and confidential information before sharing. Check your organization's data and retention rules as well, and never include credentials in prompts.