Loading patterns…
Agentic Context Engineering (Evolving Playbook)(ACE)
Treats an agent's context as an evolving playbook of concrete strategies that is improved by three roles: a Generator that produces reasoning traces on real tasks, a Reflector that extracts lessons from what succeeded or failed, and a Curator that merges those lessons into the playbook as compact structured delta updates. Because updates are incremental appends and edits rather than full rewrites, the playbook accumulates domain detail instead of suffering context collapse and brevity bias, where iterative rewriting erodes hard-won specifics. It needs no labeled supervision and applies both offline as an improved system prompt and online as agent memory, with the ICLR 2026 ACE paper reporting roughly +10.6% on agent tasks and +8.6% on finance. Distinct from `prompt-optimization`: DSPy and GEPA optimize a prompt artifact against a metric, whereas ACE grows a natural-language playbook through delta curation with an explicit anti-collapse mechanism spanning prompt and memory, and unlike `skill-library` (executable skills) or `self-improving-systems` (governed prompt and tool edits behind approval gates) the improving artifact is a curated natural-language playbook.
In 30 seconds
- What
- Divides an agent into Generator, Reflector, and Curator roles that build a natural-language playbook of strategies through incremental delta updates rather than full rewrites, preventing context collapse.
- When to use
- Multi-step domain tasks where the agent must accumulate and retain specific lessons from execution traces without labeled data, and where prompt brevity erodes hard-won detail.
- Watch out
- Curator must distinguish signal from noise; poor delta selection pollutes the playbook faster than good curation can clean it, degrading performance.
Ask the AI expert about this pattern
Opens the assistant with your question prefilled. You review it before sending.
Agentic Context Engineering (Evolving Playbook): Overview
Treats an agent's context as an evolving playbook of concrete strategies that is improved by three roles: a Generator that produces reasoning traces on real tasks, a Reflector that extracts lessons from what succeeded or failed, and a Curator that merges those lessons into the playbook as compact structured delta updates. Because updates are incremental appends and edits rather than full rewrites, the playbook accumulates domain detail instead of suffering context collapse and brevity bias, where iterative rewriting erodes hard-won specifics. It needs no labeled supervision and applies both offline as an improved system prompt and online as agent memory, with the ICLR 2026 ACE paper reporting roughly +10.6% on agent tasks and +8.6% on finance. Distinct from `prompt-optimization`: DSPy and GEPA optimize a prompt artifact against a metric, whereas ACE grows a natural-language playbook through delta curation with an explicit anti-collapse mechanism spanning prompt and memory, and unlike `skill-library` (executable skills) or `self-improving-systems` (governed prompt and tool edits behind approval gates) the improving artifact is a curated natural-language playbook.
- Generator, Reflector, Curator role division
- Context as an evolving strategy playbook
- Compact structured delta updates (append and edit, not full rewrite)
- Anti-collapse mechanism against context collapse and brevity bias
- Offline (system prompt) and online (agent memory) adaptation
- Self-improvement from execution traces without labeled supervision
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.
References
The papers, specifications, and repositories this pattern is based on.
From the engineer behind this catalog
Get your agent architecture reviewed
This page documents one pattern. Your system runs dozens, and most failures live in how they fit together. Have the whole design reviewed against the 288 patterns in this catalog: architecture, reliability, evaluation and cost, every finding mapped to the pattern that fixes it.
€750 instead of €1,500, one week, written report and walkthrough call, until 30 September