Anthropic engineers Priyanka and Gabe introduce Claude Managed Agents, explaining their evolution from a monolithic “pet” system to a modular architecture that decouples the agent’s “brain” from its “hands” to enhance scalability, security, and flexibility. They also guide attendees through creating simple agents on the Claude platform, emphasizing interactive learning and practical application with secure sandbox environments.
The presentation begins with introductions from Priyanka and Gabe, engineers at Anthropic who lead the development of Claude Managed Agents. They outline the agenda, which includes a history of building and scaling managed agents, lessons learned, and a practical workshop where attendees will create simple and more complex agents using the Claude platform. The session is designed to be interactive, with opportunities for questions throughout the two-hour seminar.
Priyanka explains the fundamental concept of an agent as a model operating in a loop with tools, making decisions and performing actions until a task is complete. She introduces the idea of a “harness,” a layer that wraps around the model to manage tool calls and context, compensating for model limitations. The importance of sandbox environments is emphasized, where agents can safely run and test code without risking security or data integrity, highlighting the need for isolation and limited attack surfaces.
The team shares insights from their initial architecture, which placed all components—session logs, harness, and sandbox execution—within a single container. While this simplified local operations, it created challenges such as difficult debugging, data privacy concerns, and inflexibility in network configurations. This setup was likened to a “pet” system, where each container was unique and required manual care, contrasting with more scalable “livestock” systems.
To address these issues, Anthropic redesigned the architecture to separate the “brain” (the agent framework and model) from the “hands” (sandbox environments and tools). This modular approach allows components to be interchangeable and independently restartable, improving fault tolerance and flexibility. They also enhanced security by ensuring sensitive credentials never enter the sandbox directly, instead using proxies to append tokens securely, mitigating risks like prompt injection attacks.
The practical portion introduces attendees to the Claude platform, where they can create environments, agents, and sessions without coding initially. Priyanka demonstrates setting up a simple “joker” agent that tells jokes, explaining core APIs and event flows such as user messages, agent responses, and custom tool usage. The session concludes with instructions on accessing API credits and navigating the platform console, preparing participants for the hands-on coding segment led by Gabe.