LangChain

Harrison Chase leads LangChain, whose sector is AI frameworks and agents. The company built LangSmith for tracing and debugging, alongside LangGraph Platform and deepagents. Its July 2025 guide How to Build an Agent sets out six steps around an email agent example: define the job with concrete examples, design the operating procedure, build an MVP focused on the core LLM reasoning tasks, connect and orchestrate the data sources, test and iterate, then deploy, scale and refine. The stated message is to start small, stay user-centred, and keep refining.

The later writing moves from method to a claim about where agent behaviour actually lives. In January 2026 Chase argued that traces document an AI application the way code documents conventional software: developers orchestrate the LLM calls, but the decisions (which tool to call, how to reason through the problem, when to stop, what to prioritise) happen inside the model at runtime. Debugging, testing, profiling and monitoring therefore shift from operating on code to operating on traces. In March 2026 Vivek Trivedy formalised the counterpart, Agent = Model + Harness, the harness being every piece of code, configuration and execution logic that is not the model itself: system prompts, tools and MCP, embedded infrastructure, orchestration logic, deterministic hooks.

The evidence LangChain offers for the harness mattering is a benchmark move, from the Top 30 to the Top 5 on Terminal Bench 2.0 by changing only the harness. Prototyping an agent with LangChain is described as relatively simple; the enterprise requirements of governance, cost control and compliance sit outside that ease.