Harrison Chase (해리슨 체이스), LangChain co-founder and chief executive officer, cited “harness engineering” as a core challenge in developing AI agents.

A harness is an execution environment designed to let an AI model loop, call tools and carry out long-running tasks.

Speaking on a VentureBeat podcast, Chase said, “Harness engineering is an extension of context engineering.” He said the recent trend is to hand more control over context engineering to the model itself. That means letting the model decide what it will and will not look at.

He said in the past it was difficult to run loops because model performance was insufficient. Citing AutoGPT as an example, Chase said it had the same structure as today’s top agents, but disappeared quickly because models at the time were not at a level to run loops reliably.

To address the issue, LangChain released a general-purpose harness called Deep Agents. Built on LangChain and LangGraph, Deep Agents provide planning, a virtual file system, context and token management, code execution, and skills and memory functions.

Work can be delegated to sub-agents, which operate in parallel based on different tools and settings. LangChain said sub-agent work is separated from the main agent’s context, and results from large-scale tasks can be compressed into a single output to improve token efficiency.

Chase stressed that for an agent to track progress and maintain consistency even across 200-step tasks, it needs a structure that proceeds while recording its thoughts.