Agent = Model + Harness

An agent's behaviour comes from the model weights plus all the non-weight runtime code around them (loop, tools, context, safety, orchestration, extensions), and that harness can be engineered independently of the model.

Evidence: moderate

When an LLM agent misbehaves, the reflex is to blame the model or reach for a bigger one. That framing hides most of what is actually running: the loop that calls the model, the tools it can invoke, how its context is assembled and trimmed, what it is allowed to do, and how work is split across sub-agents. Harness engineering starts by naming all of that as a separate artefact that can be designed, tested and changed on its own.

By 2026 the research and practitioner literature converges on the shorthand Agent = Model + Harness. A Wavestone source-code study of eleven harnesses (a July 2026 preprint) defines the harness as the runtime that couples a large language model (LLM) to the world: its loop, tools, context and safety controls. Viv Trivedy of LangChain draws the same line, treating the harness as all code, configuration and execution logic that is not the model, although his post was read only through a secondary summary. Earendil's Pi Durable post defines a harness as storage plus the machinery that runs one or more LLM conversations, together with their tools and execution environments. Microsoft's Agent Framework documentation describes it as the scaffolding that drives model and tool calls, manages state and context, applies approval policies and keeps a multistep task progressing. The framing descends from SWE-agent's agent-computer interface.

The name itself was popularised in February 2026, almost simultaneously, by Mitchell Hashimoto and by OpenAI's Ryan Lopopolo, and then formalised by Birgitta Böckeler. That timeline comes from secondary write-ups, and the exact ordering is unconfirmed.

The practical consequence is diagnostic: a fault in an agent system is at least as likely to sit in the harness as in the model, and the harness is the part a team can actually change. Inspect the loop, tools, context and permissions before concluding the model is at fault. The decomposition is refined in seven harness subsystems and inner and outer harness. How much the harness matters empirically is covered in harness as a performance lever, and how much of it to build is argued in thin vs thick harness.

Sources

  • Barbaste, Darrigol, Vu, Wiltberger (Wavestone AI Lab), arXiv 2609.00006, July 2026 preprint, §2.1 — https://arxiv.org/html/2609.00006v1
  • Viv Trivedy, LangChain, 10 Mar 2026, read via secondary summary — https://businessdatasolutions.github.io/ai-wiki/sources/2026-03-10-trivedy-langchain-anatomy-of-an-agent-harness
  • Earendil Engineering, 1 Oct 2026, harness definition — https://earendil.com/posts/pi-durable/
  • Microsoft Learn, Agent Framework concepts — https://learn.microsoft.com/en-us/agent-framework/concepts/harness