agentic-context-engineering
ACE (Agentic Context Engineering) — Zhang et al. 2025, surveyed in 2026-07-07-harness-engineering-self-improvement. Automates context-engineering by treating context as an evolving playbook of bullet points (identifier + description) rather than an ever-lengthening prompt blob.
The three components
| Component | Job |
|---|---|
| Generator | Produces task trajectories, referencing playbook bullets |
| Reflector | Distills insights from successful and failed trajectories |
| Curator | Updates the structured context with incremental, itemized entries |
The load-bearing design choice: the curator never rewrites the full context. It outputs structured bullet deltas merged by deterministic logic — preventing context collapse and brevity bias that plague iterative full-prompt rewrites. Items are refined and deduplicated periodically.
Position in the wiki
- One rung on the recursive-self-improvement ladder: prompts → structured context (ACE) → workflow → harness code → optimizer code.
- Benchmark anchor: meta-harness reports +7.7 points over ACE on online text classification while using 1/4 the context tokens — ACE is the named baseline that harness-code-level optimization is measured against.
- Contrast with MCE: ACE enforces heuristic rules for context structure (the playbook format); MCE removes the heuristics and evolves free-form skills instead.