Anthropic separates model reasoning from code execution for better agents

The architecture of AI agents

Anthropic is refining how it builds autonomous systems by distinguishing between an agent's 'brain' and its 'hands.' In this mental model, the brain represents the session log and the overarching instructions, while the hands represent the sandbox where code is actually run. By separating these two components, the company can manage complex tasks more effectively.

This distinction allows the session harness to persist even when the execution environment is temporary. If an agent attempts to run a piece of code that fails, the sandbox can be destroyed and recreated without ending the entire conversation. This prevents a single error from wiping out the progress of a long-running task.

The team describes this as a way to build higher-level behaviors out of basic primitives. For example, a feature like 'dreaming' is implemented by having one agent inspect the work of another and write memories back to the session. This modular approach allows for more complex workflows without needing to rebuild the core system for every new feature.

Improving security and latency

Separating the brain from the hands provides a significant advantage for security. Anthropic's architecture ensures that generated code runs in a sandbox that never receives a service credential. Because the credentials are held separately and attached through a proxy, there is no secret for the code to read if it tries to look around the system.

This design also helps manage where sandbox traffic can go by applying specific network controls. By isolating the execution environment from the core session, Anthropic can create a more contained space for untrusted code. This prevents the model from potentially accessing parts of the system it should not be able to reach.

From a performance standpoint, this separation helps lower latency for users. An earlier approach required starting the entire environment and agent before the first response, which added unnecessary work. By decoupling the pieces, the platform can provide faster responses while maintaining the necessary isolation for safe execution.

The path to production scale

Anthropic is moving toward a model where the base harness remains relatively simple for developers. The goal is to let users experiment with their own verifiers, parallel agents, and other custom strategies on top of the core platform. This allows for flexibility while ensuring the underlying infrastructure remains stable and reliable.

The company is also focusing on how to evaluate these agents effectively. Improving an agent harness resembles a research process where developers must pick a goal, measure behavior, and inspect failures. Anthropic is increasingly using models to help build and analyze these evaluations while maintaining human judgment over the results.

As the company prepares for its next phase, the focus is shifting toward making these systems more reliable for real-world jobs. The team is looking at how models can help each other remember information and handle multi-agent workflows. These improvements are designed to move the technology from simple chatbots to more capable virtual collaborators.

Efficiency and business growth

These technical choices are supported by a significant turnaround in the company's financial efficiency. As latent.space reported, Anthropic saw a move from a 126% loss to a 19% GAAP profit in the last year. This turnaround is attributed to two core loops: the improvement loop, where models lead to more use cases and feedback, and the investment loop, where revenue fuels more capabilities.

The company has also seen a massive increase in revenue per employee, growing nearly five times in a single year. This suggests that the team is successfully using AI to scale their own internal operations and research. These efficiencies are critical as the company positions itself as a platform for developers to build at scale.

The Claude Platform is now being used by over 50% of customers who generate at least $100,000 in revenue. This indicates a shift from providing just a model to providing a comprehensive development environment. By offering primitives that allow customers to build and deploy their own applications, Anthropic is establishing itself as a foundational layer for agentic workflows.

What comes next for developers

The roadmap for the Claude Platform emphasizes memory and learning as key areas of development. However, the team is currently prioritizing safety as model launches become more complex. They are working to ensure that as models get more capable, the decisions around safe deployment and access remain manageable for both the company and the users.

Developers can expect to see more tools for managing long-running agents that need to maintain progress across large context windows. The focus remains on building a platform that can handle autonomous work, judgment, and self-correction. This involves creating a system where agents can own a goal and work alongside humans rather than just performing single tasks.

While specific features are still being finalized, the underlying architecture is set. The move toward a 'brain and hands' model provides a blueprint for anyone designing agent runtimes. It highlights the importance of asking which parts of a system need to survive, which parts execute untrusted work, and which secrets must remain isolated.

Technical takeaways

The primary takeaway for engineers is the importance of modularity in agent design. By decoupling the state of a conversation from the environment of execution, you create a system that is inherently more resilient to errors. This is a fundamental shift from 'monolithic' agents where one crash ends the entire session.

Security practitioners should note the proxy-based credential system. This is a best practice for any AI system that interacts with external APIs or databases. It ensures that even if a model is prompted to perform a malicious action, it lacks the 'keys' to the kingdom within its execution sandbox.

Finally, the business side of the story shows that efficiency is becoming as important as raw capability. Anthropic's ability to turn a profit while scaling rapidly suggests that the 'improvement loop' is working. For the reader, this means the tools they use are becoming more sophisticated and more cost-effective simultaneously.

Implementation details

The implementation of these features will vary depending on the specific product being used. However, the core questions remain the same for any developer building with Claude. You must decide which parts of your application need to persist and which parts are destined to be ephemeral and potentially fail.

Anthropic's approach to 'dreaming' is a great example of how to build complex features using simple primitives. By having one agent inspect another's work, they create a sense of memory without needing to invent a new type of architecture. This is a scalable way to add intelligence to a system without increasing the complexity of the underlying code.

As the company moves toward its IPO, these architectural choices will likely become the standard for how they ship new models. The goal is to provide a platform that feels like a set of building blocks. This allows developers to choose how much control they want over the 'brain' versus the 'hands' of their specific AI applications.

Future of agentic workflows

The move toward 'virtual collaborators' suggests a future where AI doesn't just answer questions but completes multi-step projects. This requires the kind of reliability that the brain-and-hands architecture provides. It allows the AI to work on a task for hours or days without needing constant human intervention to restart the process.

We should also watch how Anthropic handles the 'distillation' of their models. They are currently working on protecting their model reasoning from being extracted by other models. This is a key part of their strategy to maintain the value of their high-quality outputs in a competitive market.

Ultimately, the success of these new architectures will be measured by how well they handle real-world edge cases. As more customers move to the Claude Platform, the feedback from these diverse environments will continue to shape the next generation of models. The focus remains on building a system that is both powerful enough to be useful and safe enough to be trusted.

Summary of key changes

The launch of Opus 5.5 marks a transition toward a more production-ready platform. It isn't just about a bigger model; it's about a smarter way to ship that model. By focusing on cost savings, safety, and architectural separation, Anthropic is trying to solve the practical problems of deploying AI at scale.

For the smart reader, the most important change is the shift toward agentic primitives. Instead of just asking a model to write a script, you are building a system where the model can reason about the script, run it in a safe box, and learn from the result. This is the foundation for the next wave of AI productivity tools.

The company's ability to turn a profit while doing this is a major signal for the industry. It proves that high-level research and engineering efficiency can go hand-in-hand. As Anthropic continues to iterate on these loops, the gap between 'cool demo' and 'reliable tool' will continue to shrink.

The core of the platform

At its heart, the Claude Platform is about giving developers the same tools Anthropic uses internally. This 'diffusion' of technology ensures that the best practices for safety and performance are available to everyone. It levels the playing field for smaller companies to build high-quality agents.

The distinction between the brain and the hands is a perfect metaphor for this. The brain is the intelligence provided by the model, and the hands are the tools provided by the platform. By keeping them separate, Anthropic ensures that the intelligence remains consistent while the tools can be swapped, updated, or secured independently.

This modularity is what will allow for the next generation of AI applications. Whether it's a coding assistant, a research agent, or a customer service bot, the underlying architecture remains the same. It is a robust, scalable, and secure way to bring the power of Claude into the real world.

Final thoughts on architecture

The engineering work behind Opus 5.5 shows a deep understanding of the limitations of current AI. By addressing verbosity, cost, and safety through architecture rather than just training, Anthropic is building a more sustainable product. This is a sophisticated way to handle the 'pressure' of making models more capable while keeping them safe.

The fact that they are using models to help build their own evaluations is another sign of the 'improvement loop' in action. It allows them to scale their testing capabilities at a pace that would be impossible for humans alone. This creates a virtuous cycle of better models, better tests, and better products.

For anyone looking to build with Claude, these insights are a roadmap. They suggest that the most successful applications will be those that carefully manage the boundaries between reasoning and action. By following the 'brain and hands' model, developers can build systems that are not only smarter but also more reliable and secure.

Next steps for the platform

As Anthropic moves forward, the focus will likely remain on refining these agentic primitives. We can expect to see more features related to memory, learning, and multi-agent collaboration. These are the building blocks for the 'virtual collaborators' the company is aiming to create.

The interaction between the model and the sandbox will continue to be a key area of innovation. As new types of workloads emerge, Anthropic will need to adapt its isolation and network controls. This ongoing work ensures that the platform remains at the cutting edge of both capability and safety.

The ultimate goal is a platform where the complexity of the AI is hidden behind a simple, reliable interface. By mastering the separation of the brain from the hands, Anthropic is laying the groundwork for a future where AI agents are a standard part of every workflow. It is a move from theory to practice that defines the current state of the art.

Impact on the AI landscape

Anthropic's approach sets a standard for how other labs might handle agentic workflows. The focus on 'revenue per employee' and 'revenue per dollar of compute' shows that the industry is maturing. It is no longer just about who has the biggest model, but who can run it most efficiently and safely.

The move toward a platform-centric model also changes the competitive landscape. It means that the value is moving from the model itself to the ecosystem around it. Developers who can leverage these primitives will be able to build more sophisticated and reliable applications than those who try to build everything from scratch.

As the company prepares for its IPO, these architectural and financial milestones provide a clear picture of its trajectory. Anthropic is positioning itself as a leader in the practical application of AI. By focusing on the 'hands' as much as the 'brain,' they are making Claude a tool that can actually do work in the real world.

Summary of the brain and hands

The 'brain and hands' model is a powerful way to think about the future of AI. It acknowledges that intelligence alone is not enough; you also need a safe and reliable way to act on that intelligence. By separating these two functions, Anthropic creates a system that is more robust, secure, and easier to scale.

This architecture solves several problems at once, from preventing session crashes to protecting sensitive credentials. It also allows for a more modular way of building features, like memory and dreaming. It is a sophisticated engineering solution to the complex problems of agentic AI.

For the curious reader, this story highlights the move toward practical AI. It's not just about what a model can say, but what it can do. By mastering the distinction between the brain and the hands, Anthropic is building the foundation for a new era of AI-powered work.

Conclusion on the launch

The launch of Opus 5.5 is a milestone for both Anthropic and the broader AI industry. It demonstrates how a company can achieve massive growth and profitability by focusing on efficiency and architectural clarity. The move to a platform-centric model is a key part of this strategy.

By providing developers with the primitives to build their own agents, Anthropic is empowering a new wave of innovation. The 'brain and hands' architecture ensures that this innovation is built on a secure and reliable foundation. It is a clear signal of where the technology is headed.

As we watch the next few months, the success of these new architectures will be the ultimate test. But for now, Anthropic has provided a compelling blueprint for the future of AI. It is a move toward a world where AI is not just a chatbot, but a capable, reliable, and safe collaborator.

Final architectural notes

The technical details of the sandbox and proxy system are a masterclass in secure AI design. By ensuring that credentials are never in the same space as the generated code, Anthropic addresses one of the biggest fears in AI deployment. This is a practical solution that works in the real world.

The focus on latency also shows that Anthropic understands the needs of enterprise users. A model that is smart but slow is of limited use for many tasks. By optimizing the way the agent environment starts and runs, they are making Claude a more viable tool for high-speed workflows.

In summary, the separation of the brain from the hands is a fundamental design choice. It impacts everything from security to performance to the way new features are built. It is a key reason why Anthropic is currently leading the way in the development of capable and reliable AI agents.

The future of Claude Platform

The Claude Platform is evolving into a comprehensive environment for agentic development. It is moving beyond simple API calls to a system where developers can manage complex, multi-step workflows. This is the 'diffusion' of Claude that the company highlighted in its recent filings.

By providing direct access to primitives, Anthropic is allowing for a high degree of customization. Developers can choose how to handle memory, how to verify outputs, and how to manage parallel agents. This flexibility is essential for building applications that meet specific user needs.

As the platform matures, we can expect to see more 'managed' features that simplify these complex tasks. The goal is to make it as easy as possible for anyone to build a reliable agent. The 'brain and hands' architecture is the foundation that makes this goal achievable.

Final summary of the story

Anthropic is successfully navigating the transition from a model lab to a platform provider. By focusing on architectural separation, they are solving the core problems of agent reliability and security. This is a critical step for the widespread adoption of AI agents in the workplace.

The financial success of the company further validates this strategy. By increasing revenue per employee and turning a profit, Anthropic is proving that efficient engineering is the key to sustainable growth. They are building a business that can support the next generation of AI technology.

For the reader, the most