Agentic Context Engineering (ACE): Evolving Contexts for Self-Improving LMs

peer-reviewed paper · source date 2026-01-30 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog

Rewriting a whole system prompt on each optimization step introduces brevity bias: the rewrite compresses away details it does not currently see a use for.
Repeated rewrites cause context collapse, where accumulated knowledge degrades rather than compounds.
Adapting a system to a new domain by retraining is slow and expensive relative to changing what it reads.

Treats the context / system prompt as an evolving "playbook" rather than a static string.
A Generator–Reflector–Curator loop applies incremental delta edits instead of full rewrites, which is what avoids brevity bias and collapse.
Reported gains: +10.6% on agents, +8.6% on finance, and 86.9% lower adaptation latency, matching a top AppWorld agent using a smaller model.
Stanford / SambaNova / UC Berkeley; ICLR 2026.

ACE is the bridge between context engineering and self-improvement: the artifact being optimized is agent memory that happens to live in the prompt.
Delta-edit-not-rewrite is the practical rule to take away, and it applies to any accumulating instruction file.
Named explicitly by Lil'Log's harness-engineering synthesis as a core building block.
Caveat: the original preprint predates the window; in-window status rests on ICLR acceptance and continued development. Gains are author-reported on specific benchmarks, and context-as-playbook methods share the self-feedback drift risk that SkillLearnBench and MemoryArena document.

Comments

No comments yet.