Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models
by Stanford University, SambaNova Systems & UC Berkeley
Names context collapse — and fixes it with delta updates instead of rewriting the prompt.
Overview
This ICLR 2026 paper from Stanford, SambaNova Systems and UC Berkeley proposes ACE (Agentic Context Engineering), a framework that improves LLM applications by evolving the context rather than updating weights. The authors name two failure modes in existing context-adaptation methods: brevity bias, where iterative summarisation drops domain-specific insight in favour of terse instructions, and context collapse, where repeated monolithic rewriting erodes accumulated detail over time. ACE instead treats the context as an evolving playbook and splits maintenance across three agentic roles — a Generator that produces trajectories, a Reflector that extracts lessons from execution feedback, and a Curator that merges those lessons back into the playbook. Crucially, the Curator applies structured incremental delta updates plus a grow-and-refine step rather than rewriting the whole context, which is what prevents collapse and lets the method scale with long-context models. The paper evaluates ACE both offline, optimising system prompts, and online, as agent memory, reporting +10.6% on agent benchmarks and +8.6% on finance benchmarks against strong baselines, matching top production-level agents while using smaller open-source models, and cutting adaptation latency by 86.9%. Notably it requires no labelled supervision, relying on execution feedback instead. First submitted October 2025 and revised through March 2026; 32 pages, filed under cs.LG, cs.AI and cs.CL.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~90 min read (32 pages)
- Provider
- Stanford University, SambaNova Systems & UC Berkeley
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Why iterative context summarisation causes brevity bias and context collapse
- ✓How the Generator, Reflector and Curator roles divide context maintenance
- ✓Applying structured incremental delta updates instead of monolithic context rewrites
- ✓How the grow-and-refine step keeps a playbook detailed without unbounded growth
- ✓Using execution feedback as a training signal without any labelled supervision
- ✓Reading the offline system-prompt and online agent-memory evaluation setups
- ✓What a +10.6% agent-benchmark gain and 86.9% latency reduction actually measure
Highlights
- •Published at ICLR 2026, and it named context collapse — a failure mode the field now cites
- •Covers both offline system-prompt optimisation and online agent memory in one framework
- •Reports an 86.9% adaptation-latency reduction, not only an accuracy number
- •Authored across Stanford, SambaNova Systems and UC Berkeley, including James Zou and Kunle Olukotun
- •Directly applicable to agent-memory design without any fine-tuning infrastructure
Who It’s For
Best For
- ✓Engineers designing agent memory or self-improving prompt systems
- ✓Teams hitting quality ceilings with hand-maintained system prompts
- ✓Researchers working on context adaptation as an alternative to fine-tuning
Prerequisites
- •Familiarity with LLM agents, system prompts and in-context learning
- •Comfort reading an empirical ML paper with benchmark tables
FAQ
What is Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models?
The ICLR 2026 paper introducing ACE, a framework that improves LLM agents by evolving their context rather than their weights. It identifies brevity bias and context collapse as the reasons prompt-rewriting loops degrade over time, and replaces monolithic rewrites with incremental delta updates curated by three cooperating agent roles. For engineers designing agent memory or self-improving prompts.
Is Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models free?
Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models is free to access.
What level is Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models for?
Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models is aimed at a advanced audience. Recommended background: Familiarity with LLM agents, system prompts and in-context learning, Comfort reading an empirical ML paper with benchmark tables.
How long does Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models take?
Expect roughly ~90 min read (32 pages). Most learners work through it at their own pace.
What will I learn from Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models?
You'll learn: Why iterative context summarisation causes brevity bias and context collapse; How the Generator, Reflector and Curator roles divide context maintenance; Applying structured incremental delta updates instead of monolithic context rewrites; How the grow-and-refine step keeps a playbook detailed without unbounded growth; Using execution feedback as a training signal without any labelled supervision; Reading the offline system-prompt and online agent-memory evaluation setups; What a +10.6% agent-benchmark gain and 86.9% latency reduction actually measure.
Topics
Sources
This page was written from 2 sources, 1 on domains other than arxiv.org.