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

There are two ways to make an LLM system better after training: change its weights, or change its context, meaning the instructions, examples, memory and evidence it sees. Context changes are cheap, readable and reversible, so most production improvement happens there. The question is how to accumulate what the system learns without someone hand-editing a prompt forever.

The authors identify two failure modes in existing automated approaches. Brevity bias: prompt optimizers reward short, general instructions and drop the specific tactics, tool quirks and failure modes that agents need. Context collapse: methods that ask an LLM to rewrite the whole accumulated context at every step tend to compress it. In one case study a context of 18,282 tokens was rewritten to 122 tokens in one step, and accuracy fell from 66.7 to 57.1, below the 63.7 the agent scored with no adaptation at all.

ACE (Agentic Context Engineering) treats context as a growing playbook of itemized bullets. A Generator attempts tasks with the current playbook. A Reflector reads what happened and extracts concrete lessons. A Curator turns those into small additions, a "delta", which are merged into the playbook without rewriting it. Each bullet carries counters for how often it helped or hurt, duplicates are merged by embedding similarity, and the whole thing grows and gets refined.

On the AppWorld agent benchmark and two financial reasoning benchmarks, ACE beat strong baselines by 10.6% and 8.6% on average, cut adaptation latency by 86.9% on average, and, using an open model, roughly matched the top production agent on the AppWorld leaderboard.

Every deployment generates lessons: this API paginates at 20, that date field is UTC, refunds above a threshold need approval. Today those lessons live in an engineer's head or get pasted into a system prompt that grows until someone "cleans it up" and deletes half of them. ACE is a disciplined version of that loop.

Keep lessons itemized. A playbook of separate, scored bullets can be added to, audited and pruned one item at a time. A single prompt rewritten by a model or a person loses things silently. The 18,282-to-122 collapse is the extreme version of the cleanup every team has done.

Self-improvement needs a trustworthy signal. ACE learned well from execution feedback in the agent tasks: code ran or failed, the API call succeeded or not. On the finance tasks without ground-truth labels, it degraded; online, without labels, it scored 67.3 on FiNER against a base of 70.7. A system that learns from its own guesses learns its own mistakes. Wire adaptation to real outcomes, such as test results, API errors or human corrections, or do not turn it on.

Contexts are auditable; weights are not. The authors point out that an itemized playbook supports selective unlearning: a lesson that turns out wrong, or that encodes data that must be forgotten, can be removed, which compliance teams will ask about.