← All notes

Summary

ACE grows a context as an evolving playbook, offline (for example a system prompt) and online (for example agent memory), instead of rewriting the whole context at once. The authors report average gains of 10.6% on agent benchmarks and 8.6% on domain-specific benchmarks, up to 17.1% on AppWorld, and an 86.9% average reduction in adaptation latency. [S1]

Claims

• id: C1 claim: 'They report average gains of 10.6% on agents and 8.6% on a domain split.' source: S1 evidence: "+10.6% on agents and +8.6% on finance"
• id: C2 claim: 'They report up to 17.1% on AppWorld.' source: S1 evidence: "boosts accuracy on the AppWorld benchmark by up to 17.1%"
• id: C3 claim: 'They state that code is available.' source: S1 evidence: "Our code is available at github.com/ace-agent/ace"

Verification

• Code available: Paper: github.com/ace-agent/ace, checked as https://github.com/ace-agent/ace (HTTP 200 on 2026-11-11). Not run.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live note ao-context-evolving-playbooks. This is the paper-verification note for ACE.
• Limits: Author-reported. The paper also describes a collapse case in which a context shrank from 18,282 tokens (accuracy 66.7) to 122 tokens (accuracy 57.1), worse than a 63.7 baseline without adaptation. That is their example, not a general rate.
• License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 6 Oct 2025 (v1); last revised 29 Mar 2026 (v3). Comments: ICLR 2026; 32 pages. Sources: [S1] https://arxiv.org/abs/2510.04618, v3 (29 Mar 2026). checked_on: 2026-11-11.

When to apply

Use when asking whether an agent should accumulate a playbook from execution feedback, rather than compressing history into a shorter rewrite.

Facts

• id: F1 claim: 'The authors report average gains of 10.6% on agent benchmarks and 8.6% on domain-specific benchmarks.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'They report up to 17.1% accuracy gains on AppWorld from execution feedback without ground-truth labels.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'They report an 86.9% average reduction in adaptation latency.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2510.04618, https://arxiv.org/abs/2510.04618. License: CC BY 4.0. v3 (29 Mar 2026). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

@misc{iil_note_jf7d2mkn,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2510-04618-agentic-context-engineering/},
  note = {Permanent ID iil:note:jf7d2mkn. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2510-04618-agentic-context-engineering.md.