Most discussions about AI agents focus on the model: which one is smarter, which one is faster, which one scored highest on some benchmark. That focus is understandable, but it misses the bigger question for organizations trying to get real work done with AI.
The model is only part of what makes an agent useful. The system around it is the other part, and it is often the harder part to get right.
Agent Equals Model Plus Harness
Researchers and engineers who build AI agents professionally use a simple equation: agent = model + harness.
The model is the AI itself, the part that reads, reasons, and generates output. The harness is everything else: how the agent receives context, what tools it can use, what permissions it holds, how it handles memory across sessions, how its work gets verified, and where a human must approve before anything consequential happens.
LangChain, a company that builds infrastructure for AI applications, frames it this way: the model contains the intelligence, but the harness is what makes that intelligence useful. Raw models output text. Harnesses let models use tools, act on real data, and produce work that connects to actual business systems.
Think of it like a skilled contractor. You can hire the most capable contractor in the city, but if you hand them no blueprints, no access to the site, no materials, and no way to inspect their work as it progresses, the project will stall regardless of their skill. The harness is the blueprints, the access, the materials, and the inspection process.
Why the Harness Matters More Than You Might Think
When an AI agent fails, the instinct is often to blame the model. But harness problems are frequently the actual cause.
Consider what can go wrong without a strong harness. An agent with no memory of prior sessions may repeat work already done, or contradict decisions already made. An agent with too-broad permissions may take actions no one intended. An agent with no verification step may produce outputs that look correct but are not. An agent with no clear human approval boundary may take consequential actions faster than anyone can review them.
Thoughtworks, the global technology consultancy, describes two types of harness controls: feedforward controls, which guide the agent before it acts (context, instructions, tools), and feedback controls, which catch problems after the agent acts (tests, linters, review steps). Good harness design uses both. Deterministic checks like automated tests are especially valuable because they are fast and reliable.
This framework applies well beyond software development. A research agent needs accurate context about what has already been found. A sales operations agent needs clear constraints on what it can send and to whom. An analytics agent needs verified data connections and output validation before its numbers reach a dashboard someone will use for decisions. In each case, the model is not the limiting factor. The harness is.
Long-Running Work Needs Durable Context
One harness challenge most organizations underestimate is what happens when agent work spans multiple sessions or multiple steps.
Anthropic has published engineering notes on this problem. Agents can struggle across many context windows because each new session may start without memory of what happened before. A long-running research project, a multi-step data analysis, a content production pipeline: these all require more than a single prompt and response.
Good harnesses address this by creating durable context: progress files, structured handoffs between sessions, and clear records of what has been completed and what remains. The business translation is direct: long-running AI work needs the same documentation and handoff structure that long-running human work needs. The agent has to know where it left off.
This is not a technical nicety. It is the difference between an agent that can handle a multi-day research brief and one that can only handle a three-minute task.
What OpenAI's Internal Experiment Reveals
Earlier this year, OpenAI published an account of a team that built an internal software product using an AI coding agent called Codex. According to the article, every line of application logic, tests, documentation, and internal tooling was written by the agent. The team estimates it was completed in roughly one-tenth the time it would have taken to build by hand. That figure is self-reported and worth treating with appropriate skepticism, but the directional point stands.
What matters more than the speed claim is the structure behind it. The humans involved were not idle. They designed the environment, specified the intent, and built the feedback loops that allowed the agent to produce reliable output. The framing from the team: humans steer, agents execute.
That pattern is replicable across many kinds of organizational work, not only software. The organizations that will see real return from AI agents are the ones that invest in designing the environment the agent works within, not just in selecting a capable model.
The Business Takeaway
Agent ROI does not come only from choosing a smarter model. It comes from the system around the model: context, tools, permissions, verification, feedback, and clear human approval boundaries.
Businesses that treat AI agents as a model selection problem will keep running into the same ceiling. Those that treat it as a system design problem will start seeing agents that are actually reliable, actually auditable, and actually useful across more than a demo.
Building the harness is ongoing work: refining context as needs change, updating tool permissions as the agent's scope evolves, adding verification steps as the stakes increase. It is, in many ways, the work of building a sound working environment for a new kind of team member.
The model will keep improving. The organizations that will benefit most are the ones already learning how to build a better system around it.
Sources
- Ryan Lopopolo, OpenAI, “Harness engineering: leveraging Codex in an agent-first world”
- Martin Fowler / Thoughtworks, “Harness engineering for coding agent users”
- Anthropic, “Effective harnesses for long-running agents”
- LangChain, “The Anatomy of an Agent Harness”
- OpenAI Symphony, GitHub engineering preview
- Ryan Lopopolo, harness-engineering, GitHub