Katelyn Lesse – Evolving Claude APIs for Agents, Anthropic

Key Concepts

• Agentic Systems: Systems that can autonomously perform tasks and make decisions, often by integrating with LLMs.
• Claude Code: An agentic coding product developed by Anthropic, designed to assist with coding tasks.
• Harnessing Claude's Capabilities: Exposing and utilizing the advanced abilities of Claude through API features.
• Context Window Management: Optimizing the information within Claude's context window to maximize performance.
• Giving Claude a Computer: Providing Claude with the infrastructure and tools to execute code and operate autonomously.
• Model Context Protocol (MCP): A standardized protocol for agents to interact with external systems.
• Memory Tool: A feature that allows Claude to store and retrieve information outside its immediate context window.
• Context Editing: The ability to remove irrelevant information from Claude's context window.
• Code Execution Tool: An API feature enabling Claude to write and run code in a secure, sandboxed environment.
• Agent Skills: Folders of scripts, instructions, and resources that Claude can access and execute to perform specific tasks.

Harnessing Claude's Capabilities

Anthropic is evolving its platform to help developers build powerful agentic systems using Claude. The platform focuses on three key areas: harnessing Claude's capabilities, managing its context window, and enabling it to operate autonomously with a "computer."

1. Exposing Advanced Capabilities via API:

• Reasoning and Thinking Time: Claude's performance on complex tasks improves with more reasoning time. This is exposed as an API feature allowing developers to control how much "thinking" Claude does, with an option to set a token budget for reasoning.
  • Example: Claude Code utilizes this feature to decide whether to perform a more in-depth analysis for complex debugging or provide a quick answer for simpler queries.
• Tool Use: Claude is proficient at reliably calling tools. The API supports both built-in tools (like web search) and custom tools, which can be defined with a name, description, and input schema.
  • Example: Claude Code extensively uses tools for tasks such as reading/writing files, searching for files, and rerunning tests.

Managing Claude's Context Window

Effective context window management is crucial for maximizing Claude's performance, especially for coding agents like Claude Code, which deal with diverse information like technical designs, codebases, instructions, and tool calls.

1. Model Context Protocol (MCP):

• Introduced a year ago, MCP is a standardized protocol for agents to interact with external systems.
• Real-world Application: For Claude Code, this allows interaction with platforms like GitHub or Sentry, accessing information and tools beyond its immediate context window. This leads to better performance than agents solely relying on prompt-based context.

2. Memory Tool:

• This tool allows Claude to store context outside its window and retrieve it only when needed.
• Initial Iteration: A client-side file system where developers control their data, but Claude can intelligently store and recall relevant information.
• Example: Claude Code can store codebase patterns or Git workflow preferences in memory and retrieve them when relevant to a specific task.

3. Context Editing:

• This feature helps clear out irrelevant information from the context window to make space for more pertinent data.
• Initial Iteration: Primarily focuses on clearing old tool results, which can be large and less relevant for subsequent tasks.
• Example: Claude Code, which calls numerous tools, can clear out past file reads or other tool outputs from its context window.
• Performance Impact: Combining the memory tool with context editing resulted in a 39% performance bump in internal evaluations, highlighting the importance of relevant context.

4. Expanding Context Windows and Intelligent Management:

• Anthropic is offering larger context windows, with some models supporting up to one million tokens.
• The platform is teaching Claude to better understand its context window's capacity, allowing it to adapt its responses based on available space.

Giving Claude a Computer: Autonomous Operation

Anthropic believes in empowering Claude to operate autonomously by giving it access to a "computer" and the necessary infrastructure.

1. Code Execution Tool:

• This API feature allows Claude to write and execute code within a secure, sandboxed environment hosted on Anthropic's servers.
• Benefits: Developers do not need to manage containers or security concerns.
• Example: Claude can be instructed to write and run code, such as creating an animation, and the platform handles the execution.
• Vision: The future of agents involves autonomous operation within sandboxed environments, supported by robust infrastructure.

2. Agent Skills:

• Skills are collections of scripts, instructions, and resources that Claude can access and execute within its sandbox.
• Claude decides to use a skill based on the user's request and the skill's description.
• Combination with Tools: Skills can be combined with tools like MCP, where MCP provides access to tools and context, and skills provide the expertise to utilize them effectively.
• Example: For web design tasks, Claude can utilize a "web design skill" to build landing pages that adhere to specific design systems and patterns.
• Further Information: A talk on skills by Barry and Mahes from Anthropic is recommended for deeper insights.

Platform Evolution and Future Directions

Anthropic is continuously evolving its platform to enable developers to achieve peak performance with Claude.

1. Continued API Evolution:

• As Claude's research and capabilities advance, the API will be updated to reflect these improvements, allowing developers to stay at the forefront.

2. Enhanced Memory and Context Tools:

• The platform will introduce more sophisticated tools for Claude to manage context, including deciding what to store, retrieve, and discard.

3. Focus on Agent Infrastructure:

• Anthropic will continue to address key challenges in agent infrastructure, such as orchestration, secure environments, and sandboxing, to facilitate autonomous operation.

4. Hiring:

• Anthropic is actively hiring for roles in developer products and developer relations, encouraging individuals passionate about building with Claude to reach out.