A brand new framework from Stanford University And SambaNova addresses a critical challenge in constructing robust AI agents: context engineering. Called Agentic context engineering (ACE), the framework routinely populates and changes the context window of Large Language Model (LLM) applications, treating it as an “evolving playbook” that creates and refines strategies because the agent gains experience in its environment.
ACE is designed to beat vital limitations of other context engineering frameworks and forestall the model's context from degrading as more information is collected. Experiments show that ACE is suitable for each optimizing system requests and managing an agent's memory, outperforming other methods while being significantly more efficient.
The challenge of context engineering
Advanced AI applications using LLMs rely largely on “context adaptation” or context engineering to regulate their behavior. Instead of the costly strategy of retraining or fine-tuning the model, developers use the LLMs contextual learning skills control its behavior by modifying the prompts with specific instructions, reasoning steps, or domain-specific knowledge. This additional information is often obtained because the agent interacts with its environment and gathers recent data and experiences. The predominant goal of context engineering is to prepare this recent information in a way that improves the performance of the model and avoids confusion. This approach is becoming a key paradigm for constructing powerful, scalable and self-improving AI systems.
Context engineering offers several advantages for enterprise applications. Contexts are interpretable by each users and developers, could be updated at runtime with recent knowledge, and could be shared between different models. Context Engineering also advantages from ongoing hardware and software advances, reminiscent of: growing context windows of LLMs and efficient inference techniques reminiscent of prompt and context caching.
There are various automated context engineering techniques, but most of them have two major limitations. The first is a “shortness bias,” where fast optimization methods are inclined to favor concise, general instructions over comprehensive, detailed instructions. This can impact performance in complex domains.
The second, more significant issue is “context collapse.” When tasked with rewriting all the amassed context over and all over again, an LLM may suffer from a form of digital amnesia.
“What we call 'context collapse' happens when an AI attempts to rewrite or compress every little thing it has learned right into a single new edition of its prompt or memory,” the researchers said in written comments to VentureBeat. “Over time, this rewriting process erases vital details – reminiscent of when a document is overwritten so over and over that vital notes disappear. In customer-facing systems, this may mean that a support agent suddenly loses awareness of past interactions… resulting in erratic or inconsistent behavior.”
The researchers argue that “contexts should function not as concise summaries but as comprehensive, evolving playbooks – detailed, comprehensive, and wealthy in domain-related insights.” This approach draws on the strength of recent LLMs, which may effectively extract relevance from long and detailed contexts.
This is how Agentic Context Engineering (ACE) works.
ACE is a deep context adaptation framework designed for each offline tasks reminiscent of: System prompt optimizationand online scenarios, e.g. B. Real-time memory updates for agents. Rather than compressing information, ACE treats context like a dynamic playbook that collects and organizes strategies over time.
The framework divides the work into three specialized roles: a generator, a reflector, and a curator. This modular design is inspired by “the way in which people learn – experimenting, reflecting and consolidating – while avoiding the bottleneck of overloading a single model with all responsibilities,” the paper says.
The workflow begins with the generator, which creates reasoning paths for prompts and highlights each effective strategies and customary mistakes. The Reflector then analyzes these paths to achieve vital insights. Finally, the curator consolidates these lessons into compact updates and integrates them into the prevailing playbook.
To prevent context collapse and brevity bias, ACE integrates two vital design principles. First, incremental updates are used. The context is presented as a group of structured, itemized bullets relatively than a single block of text. This allows ACE to make granular changes and retrieve probably the most relevant information without having to rewrite all the context.
Second, ACE uses a “grow and refine” mechanism. As recent experiences are gained, recent bullet points are added to the playbook and existing ones are updated. A deduplication step commonly removes redundant entries, ensuring that the context stays comprehensive, yet relevant and compact, over time.
ACE in motion
The researchers evaluated ACE on two sorts of tasks that profit from an evolving context: agent benchmarks, which require multi-turn reasoning and power use, and domain-specific financial evaluation benchmarks, which require specialized knowledge. For high-risk industries like finance, the advantages transcend just performance. As the researchers said, the framework is “much more transparent: a compliance officer can literally read what the AI has learned since it is stored in human-readable text relatively than hidden in billions of parameters.”
The results showed that ACE has strong baselines reminiscent of: B. consistently exceeded GEPA and classic in-context learning, which achieves a mean performance increase of 10.6% on agent tasks and eight.6% on domain-specific benchmarks in each offline and online environments.
Crucially, ACE can construct effective contexts by analyzing feedback from its actions and environment, relatively than requiring manually labeled data. The researchers note that this capability is a “key ingredient for self-improving LLMs and agents.” To the general public AppWorld Benchmark for evaluating agent systems, an agent using ACE with a smaller open source model (DeepSeek V3.1) corresponded to the performance of one of the best placed, GPT 4.1 based agent on average and outperformed it on the harder test set.
The advantages for corporations are significant. “This implies that corporations do not need to depend on massive proprietary models to stay competitive,” the research team said. “You can deploy local models, protect sensitive data, and still achieve superior results by continually refining context relatively than retraining weights.”
Beyond accuracy, ACE proved to be extremely efficient. It adapts to recent tasks with a mean of 86.9% lower latency than existing methods and requires fewer steps and tokens. The researchers note that this efficiency shows that “scalable self-improvement could be achieved with each higher accuracy and lower overhead.”
For corporations concerned about inference costs, the researchers note that the longer contexts produced by ACE don’t lead to proportionately higher costs. Modern deployment infrastructures are increasingly optimized for long-context workloads with techniques reminiscent of KV cache reuse, compression, and offloading, thereby amortizing the price of processing large contexts.
Ultimately, ACE points to a future where AI systems are dynamic and continually improving. “Today, only AI engineers can update models, but context engineering opens the door for subject material experts – lawyers, analysts, doctors – to directly shape what the AI knows by editing its contextual playbook,” the researchers said. This also makes governance more practical. “Selective unlearning becomes much easier to administer: if a chunk of data is outdated or legally sensitive, it could possibly simply be removed or replaced in context without retraining the model.”
