Anthropic is adding API features that let developers control Claude's reasoning time and use built-in or custom tools.
2
Memory and context editing help agents keep relevant information available while removing old tool results from the context window.
3
Anthropic is building sandboxed computer infrastructure so Claude can write and run code, use skills, and work more autonomously.
Summary
Katelyn Lesse explains how Anthropic is evolving the Claude Developer Platform for long-running agentic systems. She groups the work into three areas: exposing Claude's capabilities through features such as extended thinking and tool use, managing the context window with MCP, memory, and context editing, and giving Claude a computer through secure code execution infrastructure. Claude Code is the running example throughout the talk. Lesse describes how it uses tools to inspect and modify code, how old tool results can be removed from context, and how skills can provide domain-specific instructions and resources. She also describes the infrastructure needed for Claude Code on web and mobile, including secure environments, container orchestration, and persistent sessions. Anthropic's direction is to give developers more control over reasoning and context while making it easier for Claude to operate inside a sandbox.
The API should expose the capabilities Claude is learning
Anthropic wants developers to get direct access to capabilities that its research team is improving. Extended thinking lets an application choose whether Claude should reason longer or answer quickly, with a token budget controlling how much thinking it can use. Lesse gives Claude Code as an example because debugging a complex system may need more reasoning, while a simple request may not. Tool use is another API feature. Developers can use Anthropic's built-in web search or define custom tools with a name, description, and input schema. Claude Code uses tools to read files, search for files, write files, and rerun tests.
Context management determines what an agent can use at each moment
Lesse says that getting the right context into the window is one of the most important parts of agent performance, especially for a coding agent. Claude Code may need technical designs, an entire codebase, instructions, and many tool calls during one session. The problem is deciding which information belongs in the window at a given time and keeping that selection useful as the session continues. Anthropic's platform work addresses this with external context through MCP, stored information through memory, and removal of information that is no longer useful.
MCP lets agents reach information and tools outside their context window
Anthropic introduced the Model Context Protocol as a standard way for agents to interact with external systems. Lesse describes possible Claude Code connections such as GitHub or Sentry. These systems can provide information and tools that are not already in the agent's context. An agent that can retrieve relevant external information can do more than one that only sees what was included through prompting. MCP therefore gives Claude a way to reach systems around the coding task instead of requiring every piece of information to remain in the active window.
Memory and context editing keep the active window focused
The memory tool lets Claude keep information outside the context window and retrieve it when needed. Its first version is a client-side file system, so the developer controls the data. Lesse suggests storing codebase patterns and Git workflow preferences there. Context editing handles the opposite operation by clearing information that should no longer be active. Anthropic's first version removes old tool results, which can be large and often become less useful later in a session. Lesse says combining memory with context editing produced a 39% performance increase on Anthropic's internal benchmark.
A larger context window still needs active editing
Anthropic is also offering million-token context windows for some models. Lesse presents the larger window as something that works together with tools that edit the window, rather than as a replacement for context management. Claude is being taught to understand how much room remains and respond differently when it has plenty of space or is close to the limit. For Claude Code, this matters because a session can involve hundreds of tool calls and many files. Keeping every past result would consume space even when those results are no longer needed.
A computer gives Claude room to work through a task
Lesse argues that developers should give Claude a computer and let it work within a defined environment. Her example is Claude Code, which can write code and run that code. Anthropic's web and mobile launch exposed the infrastructure problem: a local session can use the developer's machine, but a session started remotely needs somewhere to run after the user walks away. Anthropic had to provide a secure environment for code that had not been approved in advance, container orchestration at scale, and session persistence so users could return to completed work.
Code execution moves sandbox infrastructure into the platform
Anthropic's code execution tool allows Claude to write and run code in a secure sandboxed environment. The platform handles containers and security on Anthropic's servers, so developers do not have to build those parts themselves. Lesse uses a request to make an animation more sparkly as an example of a task where Claude needs to write code and run it to see the result. She describes the direction as allowing the model to work autonomously inside a sandbox, with the platform handling the operational boundaries.
Skills add domain knowledge to tools and external context
Agent skills are folders containing scripts, instructions, and resources that Claude can choose to run in its sandbox. Claude decides when to use a skill based on the user's request and the skill's description. Lesse explains the relationship with MCP: MCP provides access to tools and context, while skills provide the expertise needed to use them. For a landing page request, a web design skill could tell Claude to follow the team's design system and established patterns. Skills can therefore connect a general coding agent to a team's own practices.
