Inner and Outer Harness

"Harness" names both the agent runtime (loop, tools, context) and the repository and environment the agent works in (AGENTS.md, linters, tests, observability), and a team running agents usually has to engineer both.

Evidence: weak

Conversations about "the harness" often talk past each other because the word is used for two different things. A dev.to survey of five vendors' definitions finds them diverging: OpenAI stresses declarative constraints, Anthropic stresses context stability across resets, LangChain uses the model-plus-harness split, and Böckeler treats the codebase itself as part of the harness.

The two meanings coexist. The inner harness is the agent runtime: the loop, the tools and context management, as described by Trivedy at LangChain, by Anthropic's Managed Agents work and by Pi. The outer harness is the repository and environment the agent works inside: AGENTS.md, linters, tests and observability, as described by Hashimoto, by OpenAI's harness-engineering post and by Böckeler. The two are engineered differently. Inner-harness work changes how the agent thinks and acts, while outer-harness work changes what the agent can see and what pushes back on it, through guides and sensors and short steering files that act as navigation pointers.

This split is a synthesis across secondary summaries rather than a distinction any single primary source draws, so the evidence is weak. It is offered as vocabulary, not as a finding.

The vocabulary is useful whenever one agent operates on systems that are themselves agents or agent-ready repositories. An operations agent that maintains other services has its own inner harness (loop, tools, permissions) and also works through an outer harness (runbooks, checks, health endpoints) of the systems it maintains, where each failure it fixes can become a new check or document. A fault report that does not say which side it is about is usually ambiguous.

Sources

  • 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
  • Ryan Lopopolo, OpenAI, Feb 2026 — https://openai.com/index/harness-engineering/
  • Software Engineering Radio, July 2026 — https://se-radio.net/2026/07/se-radio-730-birgitta-boeckeler-on-harness-engineering-for-ai-agents/
  • kenimo49, dev.to — https://dev.to/kenimo49/harness-engineering-5-companies-5-definitions-why-everyone-disagrees-on-what-it-means-531h