An approach that lets agents learn from experience without retraining.
The approach is known as Agentic Context Engineering (ACE).
How it works:
- After each run, a “reflector” agent distils experience into a short, reusable lesson.
- Lessons are stored in a structured playbook.
- On the next run, a “curator” agent retrieves relevant lessons and injects them into the prompt.
- Lessons that help are reinforced. Those that don’t are pruned.
The base model weights do not change; the working context and playbook do.
Store → Retrieve → Reinforce/Prune
Unlike Retrieval Augmented Generation (RAG), this isn’t about retrieving facts to ground a response. It’s about retrieving strategies to shape behaviour.
ACE proposes a modular, three-step agentic loop:
Generator
This is the primary “doer” agent. It attempts to complete a given task using the current version of the playbook for guidance.
Reflector
This is the “coach” agent. After the Generator finishes (or fails), it identifies why a failure occurred or what led to a success, and it distils this into a concise, actionable lesson.
Curator
This is the “archivist” agent. It takes the lesson from the Reflector and integrates it into the playbook as a new, structured entry.
Think of it as the agent writing its own operator’s manual, one task at a time.
Is it a good idea to have an agent defining its own rules?
Clowry AI