What is agentic context engineering (ACE)?

Short answer

Agentic Context Engineering (ACE) is a method from a 2025 paper by researchers including Qizheng Zhang, James Zou and Kunle Olukotun. It improves an AI agent by letting its context evolve into a playbook through three roles, generation, reflection and curation, with small incremental updates instead of rewriting. The abstract reports gains of about 10.6% on agent benchmarks and 8.6% on finance tasks, without changing model weights.

The paper

The full title is “Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models” (arXiv 2510.04618). It was first submitted on October 6, 2025 and last revised on March 29, 2026 (version 3). The thirteen authors include Qizheng Zhang, James Zou and Kunle Olukotun.

The starting point: you can adapt a language model in two ways. Change its weights (training, fine-tuning), or change its inputs (instructions, strategies, evidence). The second is cheaper, faster and readable, and it is becoming the main way agents get better at a specific domain. But the authors point out two ways existing input-based methods fail:

Problem What happens
Brevity bias Optimizers condense the context into short, generic instructions and lose the domain detail that made it useful
Context collapse Repeatedly rewriting the whole context erodes details over time, until the agent knows less than before

How ACE works

ACE treats the context as an evolving playbook: a growing collection of strategies, pitfalls and domain knowledge. Three roles work on it:

  1. Generator. Does the task and produces a reasoning trajectory: what it tried and what happened.
  2. Reflector. Distills concrete insights from the successes and errors in that trajectory.
  3. Curator. Turns those insights into small, structured updates to the playbook, instead of rewriting the whole thing.

The incremental updates are the point. The playbook is stored as itemized bullets. The curator proposes small delta entries, and plain code, not a language model, merges them into the existing playbook. The paper calls the rest “grow-and-refine”: new bullets are appended, existing ones are updated in place, and semantic de-duplication removes redundancy. Because nothing is rewritten wholesale, detail is not lost with every round.

The paper tests this on AppWorld (agent tasks) and on the financial benchmarks FiNER and Formula, among others.

What the results say

According to the abstract, ACE improved results by about 10.6% on agent benchmarks and 8.6% on finance benchmarks compared with strong baselines. It worked without labeled data, using execution feedback. The authors also report AppWorld results in a similar range to a top production agent on the leaderboard, while treating that comparison as a reference point rather than a controlled baseline. These are the paper’s own numbers on its chosen benchmarks; whether they carry over to your tasks is something to test.

The common misunderstanding

People who search for “agentic context engineering” often expect a general guide to context engineering for agents. That is a broader topic, covered in what is context engineering. ACE is narrower and more interesting: it asks whether an agent can improve its own context, so that experience becomes reusable knowledge without retraining anything.

A second misunderstanding is that a self-improving context means an ever longer context. The curator’s job is selection as much as addition. Without it you run into context rot: more text, less reliable answers.

What it means for a team

ACE is described for one agent learning from its own runs. In a team, the same idea applies across people. Every correction a colleague gives an AI is a reflection step. Every “we don’t do it that way here” is a lesson worth curating. Today those lessons mostly stay in one person’s chat history and never reach colleagues. That is a sharing problem rather than context collapse in the paper’s sense, which is detail lost through repeated rewriting, but the cure looks similar.

The practical translation for a team:

  • keep a shared playbook of rules and lessons outside any single chat
  • add to it in small, specific entries, not periodic rewrites
  • let every tool and every person’s agent read from it
  • review what goes in, because a playbook full of wrong lessons spreads mistakes as efficiently as right ones

Where Bombos fits

Bombos applies the playbook idea to a whole team: it is a plugin that gives every AI tool the team uses one shared memory, where rules learned from corrections sit next to team knowledge and what already lives in mail, documents and repos. The AI applies those rules before you ask, and you can trace which rule or source an answer rests on. It runs in the cloud or on your own server, and one person can try it first for their own agents.

Related questions

Is ACE a product I can install?

Not as a hosted product. ACE is a research framework with an official code repository, so you can run the implementation and adapt the idea, an evolving playbook with reflection and curation, to your own agent setup.

Does ACE need labeled training data?

According to the abstract, no. It can learn from execution feedback, such as whether a task succeeded, instead of labeled examples.

How is ACE different from fine-tuning?

Fine-tuning changes the model's weights. ACE leaves the model as it is and changes what goes into its context. That makes the learned knowledge readable, editable and portable between models.

Is agentic context engineering the same as context engineering?

No. Context engineering is the broad practice of deciding what a model sees. ACE is one specific research method within it, focused on letting the context improve itself over time.

Sources

  1. arXiv 2510.04618: Agentic Context Engineering, Evolving Contexts for Self-Improving Language Models
  2. Agentic Context Engineering, full text (HTML)
  3. ACE: official code repository
  4. Anthropic: Effective context engineering for AI agents
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.