Building the future of agents with Claude

Key Concepts:

• Claude Developer Platform: Anthropic's suite of APIs, SDKs, documentation, and console experiences for building on top of Claude.
• Agents: Systems where the model has autonomy to choose and call tools, handle results, and decide on the next step.
• Scaffolding: Predefined paths or constraints imposed on the model, which can limit its potential.
• Claude Code SDK: An agentic harness built around the model, originally for coding purposes, but now a general-purpose tool for prototyping agents.
• Context Management: Techniques for managing the context window (200K tokens by default, 1M in beta) to improve model performance.
• Web Search & Web Fetch: Tools that allow the model to autonomously search the web and retrieve content.
• Agentic Memory: A tool that allows the model to take notes and learn from past experiences to improve performance over time.
• Observability: The ability to monitor and audit the behavior of agents, especially in long-running tasks.
• Unhobbling the Model: Providing the model with the necessary tools and freedom to leverage its full potential.

1. Claude Developer Platform: Evolution and Scope

• The Anthropic API has evolved into the Claude Developer Platform, encompassing APIs, SDKs, documentation, and console experiences.
• The platform serves both external customers and internal products like Claude Code.
• Key additions include prompt caching, batch API, web search, web fetch, context management, and code execution.
• The platform aims to "raise the ceiling of intelligence" using Claude.

2. Defining Agents: Autonomy and Tool Use

• An agent is defined as a system where the model has autonomy to choose tools, call them, handle results, and decide on the next step.
• Anthropic emphasizes the model's reasoning and decision-making as crucial elements of an agent.
• While predefined workflows can be useful, agentic patterns allow the model to leverage improvements in future model releases.

3. The Trend Away from Scaffolding

• As models become more intelligent, the need for scaffolding decreases.
• Excessive scaffolding can constrain the model and prevent it from fully utilizing its capabilities.
• The industry is circling back to the idea of a "wild loop" with minimal constraints.
• Anthropic aims to provide lightweight tools and frameworks that enhance the model's performance without being overly restrictive.

4. Unhobbling the Model: Providing Tools and Freedom

• The focus is on "unhobbling" the model by giving it the tools it needs and allowing it to use them effectively.
• The server-side web search and web fetch tools are examples of how providing simple tools can enable complex research tasks.
• The model autonomously decides when and how to use these tools.
• "As a developer, my creativity ends at some point... but the model... will figure out a way to go do that thing."

5. Getting Started: The Claude Code SDK

• The Claude Code SDK is recommended as a starting point for developers.
• It provides an agentic harness for running the agent loop and automating tool calling.
• Originally built for coding purposes, it has evolved into a general-purpose agentic harness.
• The SDK allows developers to prototype agents without building the entire loop from scratch.
• It removes the need for developers to implement prompt caching and tool call management.

6. Business Value and Scaled Use of Agents

• When considering agents, it's crucial to identify the business value and potential impact (e.g., saving engineering hours, reducing manual work).
• Articulating the expected outcome helps define the scope of the agent project.
• The Claude Code SDK can be used for scaled deployments, but Anthropic is working on higher-order abstractions for enterprise use cases.
• These abstractions will focus on providing the best possible outcomes and leveraging Anthropic's expertise in research and inference.

7. Observability: Auditing and Steering Agents

• Observability is crucial for auditing and steering agents, especially in long-running tasks.
• It allows developers to monitor the agent's behavior, tune prompts, and adjust tool calling strategies.
• Anthropic plans to provide observability tools within the platform.
• "If we're gonna give some level of autonomy to the system, there needs to be a way to audit it."

8. Context Management: Removing Unnecessary Information

• Managing the context window is essential for optimizing model performance.
• The platform offers features to remove older, unnecessary tool calls from the context.
• The model can declutter the prompt, allowing it to focus better.
• Tombstoning: When tool calls are removed, a note is left to inform the model that the results were present.

9. Agentic Memory: Learning from Experience

• The agentic memory tool allows the model to take notes and learn from past experiences.
• The model can review its notes when stumped to improve performance over time.
• Developers can manage the memory storage location.

10. Future Directions: Higher Abstractions, Observability, and Self-Improvement

• Anthropic is focused on providing higher-order abstractions that simplify agent development and maximize outcomes.
• These abstractions will be paired with observability tools for monitoring and auditing.
• The goal is to create a flywheel where agents continuously improve over time.
• "Get to a point where... they get these really like aha moments where over time it's getting better and better and better."

11. Giving Claude a Computer

• Anthropic is exploring ways to give Claude a "computer" with persistent storage and tools.
• The code execution tool is a first step in this direction.
• This would allow Claude to perform more complex tasks, such as image analysis and data analysis.
• "It's all about unhobbling the model... just give Claude the tools."

Synthesis/Conclusion:

The conversation highlights Anthropic's vision for the future of agents, emphasizing autonomy, tool use, and minimal scaffolding. The Claude Developer Platform provides a comprehensive set of tools and features to enable developers to build powerful and intelligent agents. Key areas of focus include simplifying agent development, maximizing model performance, providing observability, and enabling self-improvement through agentic memory. The ultimate goal is to "unhobble" the model and unlock its full potential by providing it with the necessary tools and freedom to operate effectively.