Agentic Context Engineering Evolving Contexts for Self-Improving Language Models

Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

Agentic Context Engineering

Evolving Contexts for Self-Improving Language Models

info
Introduction

Large Language Model applications increasingly rely on context adaptation rather than weight updates. Current approaches suffer from two critical limitations:

Brevity bias: Over-prioritizing concise summaries at the expense of detailed domain insights

Context collapse: Iterative rewriting erodes details over time, leading to performance drops

ACE treats contexts as evolving playbooks that accumulate, refine, and organize strategies through a modular process.

architecture
Three-Role Architecture

Generator

Produces reasoning trajectories for new queries, surfacing effective strategies and pitfalls

arrow_forward

Reflector

Critiques generated traces, distilling insights from successes and errors

arrow_forward

Curator

Synthesizes insights into structured "delta entries" and integrates them into existing context

lightbulb
Key Innovations

Incremental Delta Updates

Contexts represented as structured, itemized "bullets" with metadata and content

Small, localized edits preserve prior knowledge while accumulating new insights

Non-LLM logic for deterministic merging, de-duplication, and pruning

Grow-and-Refine Mechanism

Balances context expansion with periodic refinement

Maintains relevance and prevents unbounded growth

Enables efficient, parallel merging crucial for scalability

trending_up
Performance Results

ACE consistently outperforms strong baselines across agent and domain-specific benchmarks:

+10.6%

Agent Tasks (AppWorld)

+8.6%

Financial Analysis (FiNER + XBRL)

Matches top-ranked production-level agent on AppWorld leaderboard using smaller open-source model.

speed
Efficiency Gains

ACE achieves significant efficiency improvements compared to existing methods:

Metric
Offline vs GEPA
Online vs Dynamic Cheatsheet

Latency Reduction
82.3%
91.5%

Rollout/Token Cost Reduction
75.1%
83.6%

Adapts effectively without labeled supervision by leveraging natural execution feedback.

insights
Implications

Enables scalable, efficient, and self-improving LLM systems with low overhead

Provides interpretable contexts and lower overhead compared to fine-tuning

Offers a flexible approach for online and continuous learning

Particularly valuable for specialized domains and long-context applications

arXiv:2510.04618 | Code available at github.com/ace-agent/ace