A new framework from Stanford University And Simbanova Addresses a key challenge in building robust AI agents: context engineering. It is called Agentic context engineering (ACE), the framework automatically transforms and modifies the context of large language model (LLM) applications by treating them as an “evolving playbook” that creates and optimizes strategies for the agent to experience in its environment.
ACE is designed to overcome a key limitation of other context engineering frameworks, preventing the model from overestimating its context as it collects more information. Experiments show that ACE works to both improve system cues and agent memory and improve other methods while also being significantly more efficient.
The challenge of context engineering
Advanced AI applications that use LLMs extensively "contextual adaptation," or context engineering, to guide their behavior. Instead of the expensive process of retraining or fine-tuning the model, developers use LLM Contextual learning skills Cues with specific instructions, reasoning steps, or domain-specific knowledge to guide its behavior by modifying the input. This additional information is typically acquired as the agent interacts with its environment and gathers new data and experience. A key goal of context engineering is to organize this new information in a way that improves model performance and avoids confusion. This approach is becoming a central paradigm for building scalable, self-improving AI systems.
Context engineering has several advantages for enterprise applications. Contexts are interpretable for both users and developers, can be updated with new knowledge at runtime, and can be shared across models. Context engineering also takes advantage of ongoing hardware and software developments, e.g Growing context windows LLMS and efficient modeling techniques such as instantiation and context caching.
There are automated context engineering techniques, but most of them suffer from two main limitations. The first is a “brevity bias,” where quick-fix methods favor concise, detailed, detailed instructions. This can hurt performance in complex domains.
There is a second, more serious problem "End of context." When an LLM is tasked with repeatedly rewriting its entire collection of contexts, it can suffer from a kind of digital ammonia.
“What we call ‘decontextualization’ occurs when an AI tries to rewrite or compress everything it’s learned into a single new version of its cues or memory,” the researchers said in written comments to VentureBeat. “Over time, this writing process erases important details—such as writing a document so many times that key notes disappear. In customer-facing systems, this can mean that a support agent suddenly loses awareness of past interactions … leading to incorrect or inconsistent behavior.”
The researchers argue that “contexts should serve not as comprehensive summaries, but as comprehensive, ready-made playbooks, supported, comprehensive, and rich in domain insights.” This approach leans into the strength of the modern LLM, which can effectively remove relevance from long and detailed contexts.
How Agentic Context Engineering (ACE) Works
ACE is a comprehensive context adaptation framework designed for both offline tasks, e.g System Quick Fixand online scenarios, such as real-time memory updates for agents. Instead of compressing information, ACE treats context like a dynamic playbook that collects and organizes strategies over time.
The framework divides labor into three specialized roles: a generator, a reflector, and a curator. According to the paper, this modular design is “inspired by how humans learn—helping, reflecting, and consolidating, while avoiding the constraint of overloading a single model with all the responsibilities.”
The workflow begins with a generator, which creates reasoning paths for the input signals, highlighting both effective strategies and common errors. The researcher then analyzes these passages to draw out key lessons. Finally, the curator synthesizes these lessons into compact updates and integrates them into the existing playbook.
To prevent context elimination and brevity bias, ACE includes two key design principles. First, it uses incremental updates. Context is presented as a collection of structured, itemized bullets rather than a single block of text. This allows ACE to make granular changes and retrieve the most relevant information without rewriting the entire context.
Second, ACE uses an “incremental and correct” mechanism. As new experiences are collected, new bullets are added to the playbook and existing ones are updated. The deduplication phase regularly removes redundant entries, ensuring that the context remains comprehensive and relevant and compact over time.
Ace in action
The researchers evaluated ACE on two types of tasks that benefit from evolving context: agent benchmarks requiring multidisciplinary reasoning and tool use, and domain-specific financial analysis benchmarks demanding specialized knowledge. For high-stakes industries like finance, the benefits go beyond pure efficiency. As the researchers put it, this framework is “much more transparent: the compliance officer can literally learn the AI, because it’s stored in human-readable text instead of being hidden in billions of parameters.”
The results showed that ACE consistently performed like strong baselines GEPA and classical context learning, achieving an average performance of 10.6 percent on agent tasks in both offline and online settings and 8.6 percent on domain-specific benchmarks.
Critically, ACE can create effective contexts by analyzing feedback from its functions and environment rather than requiring manually labeled data. The researchers note that this ability is a "A key ingredient for self-improvement LLMs and agents." On the people Eporled The benchmark, designed to test agentic systems, is a small open source model (agent using ACE (Depsec-V 3.1) coped with high-level performance, Agent running GPT-4.1 Averaged and surpassed on more difficult test sets.
The path taken by the business is important. “This means that companies do not need to rely on large-scale proprietary models to remain competitive,” the research team said. “They can deploy spatial models, protect sensitive data, and achieve high-level results by continuously optimizing context rather than retraining weights.”
Beyond accuracy, ACE proved highly effective. It adapts to new tasks with an average of 86.9% lower latency than existing methods and requires fewer steps and tokens. This performance demonstrates that “scalable self-improvement can be achieved with both high accuracy and low overhead,” the researchers said.
For businesses dealing with mitigation costs, the researchers noted that the longer context created by ACE does not translate to proportionately higher costs. Modern serving infrastructures are increasingly optimized for long context loads with techniques such as KV cache reuse, compression, and offloading, which reduce the cost of dealing with extensive contexts.
Ultimately, ACE points to a future where AI systems improve dynamically and continuously. "Today, only AI engineers can update models, but context engineering has opened the door for domain experts—lawyers, analysts, doctors—to directly shape what the AI ​​knows by modifying its context book." The researchers said. It also makes governance more practical. "Selective obfuscation becomes highly tractable: if a piece of information is outdated or legally sensitive, it can be easily removed or changed in context, without training the model. ”