Novel framework from Stanford University AND SambaNew solves a key challenge in building hearty AI agents: context engineering. Called Agent context engineering (ACE) framework automatically populates and modifies the huge language model (LLM) application context window, treating it as an “evolving playbook” that creates and refines strategies as the agent gains experience in its environment.
ACE is designed to overcome key limitations of other context engineering platforms by preventing model context from degrading as more information is collected. Experiments show that ACE performs well in both system prompt optimization and agent memory management, outperforming other methods while being significantly more competent.
The challenge of contextual engineering
Advanced AI applications using LLM rely heavily on “context adaptation” or context engineering to guide their behavior. Instead of an high-priced process of retraining or re-tuning the model, developers benefit from LLM learning skills in context guide its behavior by modifying input prompts with specific instructions, reasoning steps, or domain-specific knowledge. This additional information is typically obtained as the agent interacts with its environment and gathers modern data and experiences. A key goal of context engineering is to organize modern information in a way that improves model performance and avoids model confusion. This approach is becoming the main paradigm for creating competent, scalable and self-improving artificial intelligence systems.
Context-aware engineering has several advantages for enterprise applications. Contexts are interpretable to both users and developers, can be updated with modern knowledge at runtime, and can be shared across models. Context-aware engineering also benefits from continuous advances in hardware and software, e.g growing context windows LLM and competent inference techniques such as cue and context caching.
There are many automated context engineering techniques, but most of them have two key limitations. The first is the “brevity bias,” where rapid optimization methods favor concise, general instructions over comprehensive and detailed ones. This can degrade performance in intricate domains.
The second, more sedate problem is “context collapse”. When an LLM is tasked with repeatedly rewriting all of the context it has gathered, it may suffer from a kind of digital amnesia.
“What we call ‘context collapse’ occurs when the AI tries to rewrite or compress everything it has learned into one modern version of the cue or memory,” the researchers said in written comments to VentureBeat. “Over time, this rewriting process removes critical details — for example, overwriting a document so many times that key notes disappear. In customer-facing systems, this may mean that the support worker suddenly loses awareness of past interactions… causing erratic or inconsistent behavior. “
The researchers argue that “contexts should function not as concise summaries, but as comprehensive, evolving textbooks – detailed, inclusive, and rich with field-specific insights.” This approach builds on the power of modern LLMs, which can effectively extract meaning from long and detailed contexts.
How agent-based context engineering (ACE) works.
ACE is a framework for comprehensive context adaptation, intended for both offline tasks, such as: system optimizationand online scenarios such as real-time agent memory updates. Instead of compressing information, ACE treats context as a dynamic playbook that collects and organizes strategies over time.
The structure divides the work into three specialized roles: generator, reflector and curator. According to the article, this modular design is inspired by “how people learn – by experimenting, reflecting and consolidating – while avoiding the bottleneck of overloading a single model with all responsibilities.”
The workflow starts with a generator that creates reasoning paths for the prompts, highlighting both effective strategies and common mistakes. Reflektor then analyzes these paths to draw key conclusions. Finally, the curator synthesizes these lessons into compact updates and connects them to the existing textbook.
To prevent context wrap and brevity bias, ACE incorporates two key design principles. First, it uses incremental updates. Context is represented as a collection of organized, detailed bullets rather than a single block of text. This allows ACE to make detailed changes and retrieve the most relevant information without having to rewrite the entire context.
Second, ACE uses a “grow and refine” mechanism. As new experiences are accumulated, new points are added to the manual and existing ones are updated. The de-duplication stage regularly removes redundant entries, ensuring that the context remains comprehensive, yet relevant and compact over time.
ACE in action
Researchers evaluated ACE for two types of tasks that take advantage of a changing context: agent benchmarks that require multidirectional reasoning and tool use, and domain-specific financial analysis benchmarks that require specialized knowledge. In high-stakes industries like finance, the benefits go beyond sheer performance. As the researchers said, the framework is “much clearer: a compliance officer can literally read what the AI has learned because it is stored in human-readable text rather than hidden in billions of parameters.”
The results showed that ACE consistently outperformed strong baselines such as GEPA and classic context learning, achieving an average performance increase of 10.6% on agent tasks and 8.6% on domain-specific benchmarks, both offline and online.
Most importantly, ACE can build effective contexts by analyzing feedback from its activities and environment, rather than requiring manual data marking. The researchers note that this ability is “a key component of LLM and agent self-improvement.” On public opinion The world of applications benchmark, intended for evaluating agent systems, an agent using ACE with a smaller open source model (DeepSeek-V3.1) equaled the results of the highest ranked ones, GPT-4.1 based agent average and exceeded it on the more difficult test set.
The benefits for businesses are significant. “This means companies don’t have to rely on huge, proprietary models to stay competitive,” the research team said. “They can deploy local models, protect sensitive data, and still achieve top-notch results by continuously refining context rather than retraining weights.”
Apart from accuracy, ACE proved to be very efficient. It adapts to new tasks with an average of 86.9% lower latency than existing methods and requires fewer steps and tokens. The researchers emphasize that this efficiency shows that “scalable self-improvement can be achieved with both greater accuracy and less effort.”
For companies concerned about inference costs, researchers point out that the longer contexts created by ACE do not translate into proportionally higher costs. Up-to-date service infrastructures are increasingly optimized for long-context workloads with techniques such as KV cache reuse, compression, and offloading that amortize the cost of serving long context.
Ultimately, ACE points to a future in which AI systems are lively and constantly improving. “Today, only AI engineers can update models, but contextual engineering opens the door for domain experts – lawyers, analysts, doctors – to directly shape AI’s knowledge by editing its contextual playbook,” the researchers said. This makes management more practical. “Selective unlearning becomes much easier: if information is outdated or legally sensitive, it can simply be removed or replaced in context, without having to retrain the model.”