Katelyn Lesse, engineering lead for the Claude developer platform at Anthropic, presents a three-pillar approach to maximizing the performance of agentic systems with Claude, using Claude Code as a concrete implementation example.
The first pillar consists of harnessing Claude's capabilities through customizable API features. Claude has developed a "thinking" capability where its performance improves with the amount of reasoning time allocated. The API exposes this feature with a configurable token budget, allowing developers to choose between fast responses and in-depth reasoning. Reliable tool use allows Claude to call custom tools, as Claude Code does with its many file manipulation and test execution tools.
The second pillar concerns context management. Getting the right context at the right time is critical for performance. Three major innovations support this: MCP (Model Context Protocol) enables standardized interaction with external systems such as GitHub or Sentry; the Memory tool offers a client-side file system where Claude intelligently stores information (codebase patterns, git preferences) to recall when needed; Context Editing automatically cleans up obsolete tool results, generating a 39% performance gain in internal benchmarks.
The third pillar, the boldest, advocates "giving Claude a computer". Rather than debating the complexity of agent harnesses, Anthropic provides the infrastructure enabling Claude to write and execute code autonomously. The Code Execution Tool offers a secure sandboxed environment on Anthropic's servers, managing containers and orchestration. This approach enabled the launch of Claude Code on web and mobile, solving complex orchestration and session persistence challenges.
Agent Skills enrich this autonomy by providing domain expertise. These are folders of scripts and instructions that Claude can invoke depending on context, such as automatically applying a design system when creating landing pages. This combination of Skills with MCP creates a system where Claude has access to tools, context, and the expertise to use them effectively.
Anthropic's future vision continues to evolve around these three axes: exposing new Claude capabilities via the API, improving context management with more sophisticated tools, and solving infrastructure challenges to enable genuine secure agentic autonomy at scale.