← back
Katelyn Lesse – Evolving Claude APIs for Agents, Anthropic
Takeaway
Anthropic is shipping memory + context-editing + sandboxed code execution as primitives so developers don't build agent harness infra themselves.
Summary
- Caitlyn Lesse (Anthropic Claude developer platform lead) frames the Claude API evolution around three pillars: harness Claude's capabilities, manage its context window, and give Claude a computer.
- Capabilities surface as API features: extended thinking with a token budget, plus tool use (built-in like web search and custom tools with name/description/input schema).
- Context management spans MCP for external tools, a clientside-filesystem memory tool, and context-editing that clears stale tool results; combining memory + context editing yielded a 39% bump on internal evals, with up to 1M-token windows on some models.
- 'Give Claude a computer' — Anthropic launched Claude Code on web/mobile with their code execution tool: sandboxed container orchestration handled server-side so users don't manage infra.
- Claude is being trained to be aware of how much of its context window remains and to plan accordingly.
anthropicclaudemcp
Original description
Developers are building more and more complex, long-running, agentic systems. Learn how the Anthropic team is evolving the Claude Developer Platform to enable developers to get the best outcomes from Claude.