Lance Martin

Lance Martin develops and explains AI agent systems, with particular attention to context management and designs that can adapt as models improve. At LangChain, he worked on open-source and applied projects; his Open Deep Research redesign replaced rigid parallel report writing with flexible research followed by a final writing step. In April 2026, he coauthored Anthropic’s account of Managed Agents with Gabe Cemaj and Michael Cohen, describing how the team separated agent harnesses, execution environments, and durable session history.[8][5][7]

In his June 23, 2025 article, Context Engineering for Agents, Martin organized existing approaches into four categories: write, select, compress, and isolate. He drew on Andrej Karpathy’s analogy of the model as a CPU and its context window as working memory, and Drew Breunig’s descriptions of context poisoning, distraction, confusion, and clash. Martin’s synthesis addressed a practical problem: long-running agents accumulate tool feedback that can exhaust the context window, increase cost and latency, and degrade performance.[4]

Writing saves information outside the context window; selecting brings relevant information back in. Martin distinguished scratchpads used within a task from memories retained across sessions. A scratchpad might be a file accessed through tools or a field in persistent runtime state. Selection then determines which plans, facts, examples, or instructions the model sees at each step. He also emphasized that retrieval can select tool descriptions as well as knowledge, reducing overload from large or overlapping tool collections.[4]

Compression reduces the tokens needed for a task through summarization or trimming. Isolation separates context across subagents, execution environments, or runtime-state fields. Martin used Anthropic’s multi-agent researcher as an example of assigning narrow subtasks to separate context windows, while noting the costs of additional tokens and coordination. He also described Hugging Face’s CodeAgent: code executes in a sandbox, where large objects can remain as variables while only selected results return to the model.[4]

Martin and Drew Breunig held a context-engineering meetup in July 2025. That September, Martin appeared on Latent Space, identified with LangChain, for an episode covering context engineering, memory, retrieval, and agent frameworks. In his first-person BRXND biography, he credits himself with creating many LangChain Quickstarts, LangChain’s first full course, LangChain Academy, and RAG From Scratch.[1][8]

Martin’s July 30, 2025 account traced Open Deep Research’s design to frustrations with unreliable tool calling and small context windows in 2023. In early 2024, he turned toward workflows with predefined code paths. His late-2024 orchestrator-worker system decomposed a request into report sections, researched and wrote those sections in parallel, and combined the results.[5]

As tool calling improved and MCP gained momentum, those choices became constraints: the workflow could not use the growing tool ecosystem, always imposed a section-based research plan, and sometimes produced disjoint reports. Martin moved to a multi-agent system but initially kept parallel section writing, preserving the coherence problem. His subsequent decision was concise: “I moved writing to a final step.” Agents could gather context under a flexible research plan, then a single final step could write from the collected material.[5]

Martin reported a Deep Research Bench score of 43.5 and a top-ten placement in July 2025. He credited Vadym Barda for initial evaluations, MCP support, and discussion, and Nick Huang for work on the multi-agent implementation and Deep Research Bench evaluations.[5]

Drawing on Rich Sutton’s Bitter Lesson and Hyung Won Chung’s research talk, Martin argued that engineers should understand the assumptions embedded in their applications, reconsider them as models improve, and make imposed structure easy to remove. He retained LangGraph for features such as checkpointing while favoring low-level nodes and edges that he could reconfigure.[5]

In Agent design patterns, published January 9, 2026, Martin argued that effective agent design largely comes down to context management. He highlighted filesystems for persistent context, shells for executing utilities and code, and progressive disclosure for loading detailed instructions or tool descriptions only when needed. Offloading old results to files can preserve information that summarization might lose; prompt caching can reduce the cost of repeated context. He presented these as patterns synthesized from other practitioners’ work.[6]

Martin also described experiments inspired by Cat Wu: distilling Claude Code sessions into diary entries, reflecting across those entries, and updating CLAUDE.md. He tested related skill-learning ideas with Deepagents. Looking ahead, he suggested that models might absorb some of the prompting and scaffolding used for context management in agent harnesses.[6]

The April 8, 2026 article he coauthored with Gabe Cemaj and Michael Cohen, Scaling Managed Agents: Decoupling the brain from the hands, describes a team redesign after a single-container architecture created reliability, debugging, and connectivity problems. The team separated the harness that calls Claude, the sandboxes and tools that execute actions, and the durable session log. A replacement harness could recover the event history after a crash; stored events could also be retrieved beyond the model’s current context window.[7]

The team also kept service credentials beyond the generated-code sandbox’s reach, using mechanisms including an external OAuth vault and an MCP proxy. Separating inference from immediate container provisioning let sessions begin responding before a sandbox was needed. The authors reported roughly 60% lower median time to first token and over 90% lower p95 latency in their internal comparison. They acknowledged Nodir Turakulov and Jeremy Fox for discussions, and the Agents API team and Jake Eaton for contributions.[7]

Before his LangChain work, Martin earned a Stanford PhD and spent several years as a technical lead or manager on vision systems in the self-driving industry, according to his first-person biography. His earlier doctoral work provides a concrete example of his experience building software around research data.[8][9]

In his February 2015 account, Martin named Steve Quake as his advisor and described his group’s reuse of cell-free-DNA sequencing datasets from transplant patients. The group recovered nonhuman fragments normally discarded by the processing pipeline, matched them against microbial genomes, and counted the matches. They compared the resulting signals with independent clinical tests.[9]

The group used Postgres, Pandas, and SQLAlchemy to join microbial counts with taxonomy, patient metadata, clinical results, and pipeline logs. Their application organized the data into cohorts, patients, samples, and detected microorganisms. Patient tables linked to infection views showing time series alongside clinical tests and sequencing coverage, giving clinicians and researchers ways to explore the measurements.[9]