Most conversations about AI agents open with the wrong question: which model is best? It is a natural question, because it is what the tech press repeats at every launch and at every new version of an AI model. It also explains less and less about the result your company will see in its operation.
What a harness is, in plain terms
The AI model is the engine. The harness is the rest of the car: what an AI agent receives as instruction, which tools it is allowed to call, what it keeps in memory between steps and between runs, how many times it may retry, how it checks its own work, and when it stops and hands an approval to a person (the human gate).
Two AI agents can run exactly the same model, with the same prompts, and deliver very different results, because the layer around them is different. In production, the harness is where most of the agent engineering lives — and a large share of the outcome.
The effect size, measured
Terminal-Bench 2.0 is a public battery of terminal tasks: the kind of work where an AI agent has to chain many steps without losing the thread. The numbers below help size the effect.
Treat that as an order of magnitude, not a controlled experiment: these are two separate submissions to the same benchmark. The honest reading is "holding the model constant and varying the agent harness" — not "harness X beats harness Y for your company." Different kinds of task may well suit different harnesses, inside the same company.
The next data point is more direct, because there the team changed only the harness layer:
And there is a third effect that rarely makes it into the math: the infrastructure the agent runs on.
So beyond the argument about the AI model and the execution harness, part of the result depends on the infrastructure. That is even more true when you run open weight models on infrastructure of your own.
Why this happens
The AI model supplies potential capability. The harness decides how much of that capability becomes finished work, and how the process is orchestrated. In practice, four decisions explain most of the difference:
- Instruction. What the agents know about each process, about the workflow, about what is forbidden and what is allowed, and about how results will be judged.
- Tools. What the agents can actually do — query systems and data, write records, open tickets, send messages — and within which limits and authority.
- Context. What the agents carry and remember between steps and between runs. Too much context confuses them and costs money in tokens; too little leads them to hallucinate.
- Verification. How tasks are checked before they count: a second agent reviewing, a deterministic rule, or a human gate.
Model
Potential capability: the engine.
Harness
Decides how much of that capability becomes finished work.
- 01Instruction
- 02Tools
- 03Context
- 04Verification
Infrastructure
Where the agent runs, which also changes the outcome.
Those four decisions are engineering decisions your company makes; they do not depend on the model vendor. That is why switching models rarely fixes an agent that keeps getting things wrong, and why the same engineering team can improve results without changing anything in the AI inference layer.
What changes when you design a team of AI agents
The question stops being which model is smarter and becomes which combination of model, harness, and infrastructure delivers this task within acceptable cost, time, and risk.
That has a concrete consequence: if much of the value sits in the harness layer, that layer has to be your company's, or at least portable between vendors:
An AI agent that exists only inside a vendor's product is a dependency that is hard to price — and it surfaces precisely when you need to negotiate, especially when something goes wrong.
There is also a question of human staffing:
The scarcity is not access to AI models, which are one API call away. It is people proficient in agentic AI, who can design, measure, and operate agents in business processes.
How to test this in your own company
You do not need public benchmarks to decide. The method is the same one those benchmarks use, at a smaller scale:
- Pick 20 to 50 real tasks from a process you are considering automating with AI agents, where the correct outcome is already known.
- Freeze one AI model and vary one layer at a time: instructions and flows first, then the agents' skills and tools, then verification.
- Record the same numbers for every unit of work: token cost, infrastructure cost, human intervention rate, and accuracy rate.
- Repeat with another model only at the end, once the harness layer is stable. Changing everything at once makes it impossible to know what caused what.
That set of decisions — instrumented, measured, and owned — is what AInda calls its method. The practical result is not just more effective AI agents: it is agents whose behavior you can explain, predict, and improve continuously.