Katelyn Lesse is Anthropic’s Head of Platform Engineering, leading the teams that turn Claude into dependable APIs, developer tools, and infrastructure for autonomous software. Her work addresses the practical requirements of useful agents: durable sessions, secure execution, access to external systems, and carefully managed context.
Before Anthropic, Lesse led Connect engineering at Stripe and was a senior engineering director at Betterment, where she led cash-management engineering after overseeing Betterment for Business engineering. Stripe Accounts v2 provides configurations for merchants, customers, and payment recipients. Stripe discourages indefinitely maintaining both Accounts API versions simultaneously.
At Anthropic, she initially led the Claude Developer Platform team before becoming head of platform engineering. Her organization developed Claude Managed Agents, which entered public beta in April 2026. An early design bundled an agent’s reasoning loop, execution environment, and session state inside one container, creating slow starts, fragile sessions, and security concerns. The team rebuilt the system around separate components for reasoning, sandboxed execution, and durable state. Following Anthropic’s acquisition of Stainless, Lesse also emphasized the importance of connecting agents to external services through developer infrastructure such as SDKs, APIs, and Model Context Protocol servers.
Agent infrastructure is a distributed-systems problem. Reliable agents require isolated execution, durable state, credential boundaries, orchestration, and recovery. Lesse argues that making a single agent runner the durable component creates unnecessary machinery for lifecycle management, checkpointing, and replay; separating execution from persistent state avoids that architectural trap. Her critique of agent-runner design applies that principle beyond Anthropic.
Context management determines agent performance. Agents need to retrieve relevant information, store reusable knowledge outside the immediate context window, and discard stale tool results.
Human judgment, agentic execution. Lesse assigns motivation, product taste, and customer trust to people while giving agents implementation, testing, review, and incident investigation. Her model for human-agent software teams uses self-verifying agent loops that plan, write code, test their output, and continue until they meet an explicit standard.
Claude Code illustrates what an agent platform must provide: control over model capabilities, selective context management, and a secure computer on which to work.
Katelyn Lesse and Angela Jiang show why an agent’s token budget is also an allocation problem: advice, evaluation, and reflection can outperform spending the same allowance entirely on execution.
Treat an agent budget as an allocation across jobs, not merely a quantity of execution tokens.
Control the token allowance when comparing strategies. On the reported benchmark, execution scored 76 and advising scored 89 under the same roughly 600,000-token maximum.
Measure the outcome the user can actually use. For the P&L example, anything short of a perfectly scored answer still requires correction or another run.
Claude Managed Agents supplies the concrete individual-agent layer described in the talk; a meta-harness composes roles above it, while automatic strategy construction remains a longer-term goal.
