-
Notifications
You must be signed in to change notification settings - Fork 0
Home
Evolving, self-improving context playbooks for LLM agents.
ACE is a clean, dependency-light, fully tested implementation of the ICLR 2026 paper "Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models", with first-class OpenAI Agents SDK support.
Stop re-prompting. Let your agent write its own playbook from experience.
LLM agents increasingly improve through context adaptation — editing their inputs (instructions, strategies, evidence) instead of their weights. Two common approaches fail: prompt optimizers suffer brevity bias (they collapse to short, generic instructions and discard hard-won detail), and letting an LLM rewrite the whole context every step causes context collapse (a lossy summary that craters accuracy). ACE fixes both by treating context as an evolving playbook of small, itemized bullets that accumulate, refine, and organize over time — via a modular Generator → Reflector → Curator loop with incremental delta updates and a deterministic grow-and-refine step.
| Page | What's on it |
|---|---|
| Quickstart | Install, run the headline demo (no API key), 10-line Python example |
| How It Works | The Generator/Reflector/Curator loop, deltas, grow-and-refine, offline vs. online |
| Architecture | Full diagram set, data model, module map, runtime data flow |
| OpenAI Agents SDK |
wrap_agent / ACEAgent: drop-in self-improving memory |
| Custom Tasks and Feedback | Bring your own Task and label-free feedback_fn
|
| Cookbook | 10 short, runnable recipes (7 need no API key) |
| Configuration |
ACEConfig knobs, backends, embedders, curation modes |
| Results | Reproducible offline numbers + paper benchmarks |
| FAQ | Design decisions, when ACE is (and isn't) worth it |
| Glossary | Key terms in one place |
| Prompt optimizers (GEPA, MIPRO) | Monolithic memory (full rewrite) | ACE | |
|---|---|---|---|
| Keeps domain detail | ❌ brevity bias | ✅ accumulates | |
| Survives long horizons | ❌ context collapse | ✅ incremental deltas | |
| Update cost | 🐢 full re-optimization | 🐢 full re-ingest each step | ⚡ tiny deltas, non-LLM merge |
| Works without labels | ✅ | ✅ execution feedback | |
| Interpretable / editable | ✅ inspectable bullets |
- 📦 Repository: https://github.com/rrahimi-uci/agentic-context-engineering
- 📖 Docs site: https://rrahimi-uci.github.io/agentic-context-engineering/
- 📐 Architecture (rendered): https://rrahimi-uci.github.io/agentic-context-engineering/architecture.html
- 📄 Paper: arXiv:2510.04618
This repository is an independent, open-source reproduction for research and educational use. All credit for the ACE method belongs to the original authors.
ACE — Agentic Context Engineering · MIT · Independent open-source reproduction of ICLR 2026 arXiv:2510.04618. All credit for the method belongs to the original authors.