ACE Framework (ICLR 2026): Self-Improving Agent Contexts via Delta Updates — +10.6% on Agent Benchmarks Without Fine-Tuning

arXiv / ICLR 2026·high signal

The Agentic Context Engineering (ACE) framework, accepted at ICLR 2026, treats agent contexts as evolving playbooks with three specialized components: Generator (produces reasoning traces), Reflector (extracts lessons), and Curator (synthesizes delta updates merged deterministically without LLM calls). Instead of rewriting full prompts, ACE performs localized delta edits that accumulate insights while preserving prior knowledge. Results: +10.6% on agent benchmarks, +8.6% on finance, matches top AppWorld leaderboard with a smaller open-source model.