Lance Martin

Lance Martin is a member of technical staff at Anthropic working on the Claude Platform, including Claude Managed Agents and the claude-api skill in Claude Code. Previously an early LangChain engineer, he builds systems that let increasingly capable AI agents work independently while remaining observable, secure, and responsive to feedback.

From self-driving vehicles to AI agents

Martin earned a PhD at Stanford and worked on computer vision and perception for autonomous vehicles at Uber ATG, Ike, and Nuro. At LangChain, he contributed to LangGraph, developer education, retrieval-augmented generation, and evaluation; created the first LangChain Academy agent-building course and RAG From Scratch; and built auto-evaluator, an evaluation tool for language-model question-answering systems.

His 2024 work on controllable agents positioned graph-based workflows between rigid, developer-defined chains and unpredictable open-ended agents. Explicit state and constrained branching let models choose among approved paths while supporting corrective retrieval, hallucination checks, and code-validation loops. By 2026, at Anthropic, he had extended those concerns to infrastructure for agents running asynchronously over much longer periods.

How Martin designs dependable autonomy

• Separate orchestration from execution. In long-horizon Claude systems, the agent harness coordinates isolated execution environments through a durable, append-only session. Failed workers can restart without erasing progress, credentials remain outside sandboxes, and models can revisit earlier context instead of relying solely on compressed summaries.

• Evaluate the route and independently verify results. His approach to agent trajectory evaluation examines wasted tool calls, inappropriate repetition, and other failures hidden by a correct final answer. For extended assignments, a separate verifier evaluates completed work against explicit outcomes, preventing the working agent’s accumulated assumptions from shaping its own grade. He demonstrated verifier-driven iteration using Parameter Golf, a benchmark created by OpenAI.

• Repair memory instead of merely accumulating it. His claude-diary project implements a simple memory system for Claude Code. His experiments distinguish notes recorded during execution from offline memory consolidation through dreaming, which reviews previous sessions, identifies misleading memories, and improves future decisions. He favors flexible filesystems or databases that models can organize themselves over rigid, developer-prescribed memory schemas.

• Engineer context around actual model behavior. His writing on agent design emphasizes compact toolsets, progressive disclosure, persistent filesystems, and recoverable external context. For retrieval, he separates small chunks used to locate relevant information from larger documents supplied for answer generation; long context windows alone cannot guarantee reliable recall. His analysis of Manus further examines these context engineering tradeoffs.

Martin’s essay on the bitter lesson describes the architectural shift accompanying stronger models: fewer brittle, hand-coded assumptions; more general-purpose tools, persistent state, secure boundaries, and measurable feedback. His outline of emerging agent patterns extends that thinking to organization-level agent harnesses: shared systems with independent identities, organizational context, multiple users, and the ability to surface useful information proactively.