Stop Asking If 97 Is Prime: This Is the Era of Harnesses

The prime-number test is a 2023 artifact. In 2026 the value sits in the system around the model, and that system can be measured.

The wrong question

Asking a model whether 97 is prime tells you nothing about whether it can do your work. It is prime. The model will say so. Nothing has been learned.

That test belongs to a debate that peaked in 2023. Since then, the value in enterprise AI has moved out of the sentence you type and into the system around the model.

That system is the harness: the tools a model may call, the context it receives, the contract it works under, and the checks it must pass before anything ships.

A 2023 debate built on 2022 research

The prime-number debate is mostly a 2023 story. Its techniques come from 2022 research, it peaked in mid-2023, and the field has renamed itself twice since.

When	Era	What happened
Jan 2022	Prompt craft	Chain-of-thought prompting shows that worked reasoning examples in the prompt lift math scores.
Nov 2022	Prompt craft	ChatGPT launches. Prompt wording becomes the main lever most users have.
Jul 2023	Prompt craft	Chen, Zaharia and Zou report GPT-4’s prime-number accuracy falling from 97.6% in March to 2.4% in June.
Jul 2023	Prompt craft	Narayanan and Kapoor note the test held only primes. On composite numbers, the June model did far better.
Mar 2024	Prompt craft	The revised paper uses a balanced set: 84% falls to 51%, partly because GPT-4 followed chain-of-thought instructions less.
Jun 2025	Context	Tobi Lütke and Andrej Karpathy back context engineering over prompt engineering.
Dec 2025	Context	The Linux Foundation forms the Agentic AI Foundation around MCP, AGENTS.md and goose.
Feb 2026	Harness	Mitchell Hashimoto writes about engineering the harness. OpenAI follows with its harness engineering memo (overview).
Feb 2026	Harness	LangChain lifts its coding agent from 52.8% to 66.5% on Terminal-Bench 2.0. The model stayed the same.

Bringing the prime-number test into a 2026 architecture discussion is like judging a car by revving its engine in the garage.

Why “is 97 prime?” tells you nothing

The prime test fails as an evaluation in four ways, and each one points to the harness.

It has a deterministic answer. A three-line function decides primality with 100% accuracy at near-zero cost. A good harness lets the model call that function instead of recalling the answer.
It measures a default, not a capability. Kapoor’s reading was that GPT-4 had simply shifted its default answer from yes to no. On a test set of only primes, that looks like collapse.
It is the wrong unit. Enterprises do not buy answers. They buy completed tasks: a merged change, a reconciled invoice, a closed incident, each measured end to end.
It hides the real lesson of 2023. The paper’s lasting finding was drift: the same model name behaved differently three months later. Pinned versions, regression evals and monitoring fix drift. That is harness work, not prompt work.

What a harness actually is

An agent is a model plus a harness. The model is the engine. The harness turns its output into finished, checked work.

The clearest evidence is LangChain’s February 2026 experiment. With gpt-5.2-codex held fixed, changes to prompt, tools and middleware alone added 13.7 points. Their most common failure: the agent wrote a solution, re-read its own code, judged it fine, and stopped.

A production harness has seven parts:

Tools. Code execution, search, MCP servers. The model calls a function instead of guessing.
Context. AGENTS.md, retrieved documents, memory. Curated per task, never dumped into the window.
Contract. A typed task spec. In the architecture I work on, no agent spawns without six fields: goal, success criterion, budget, depth allowance, error budget and return schema.
Verification. Tests and checks run before the agent may report done. Agents do not grade their own work.
Budgets and loop control. Time, token and retry limits. LangChain saw agents repeat the same broken approach more than 10 times.
Gates. No autonomous change to production models or infrastructure without a human gate and a documented evaluation.
Traces. Every action is recorded, so a failure becomes a fix in the harness instead of an anecdote.

Every mistake an agent makes should end as a permanent change to its harness, not as a better-worded prompt.

Measure the system, not the sentence

If the harness is the product, its outcomes are the metric. DORA’s 2025 research, with nearly 5,000 respondents, finds AI acts as an amplifier: it strengthens strong organizations and magnifies the dysfunctions of weak ones. The harness is where you steer that amplification.

Metric	What it answers
Task success rate	Share of tasks that meet the contract’s success criterion, checked by tests rather than by the agent
Cost per completed task	Tokens, compute and human review minutes per finished task, not per call
Change failure rate	How often agent-authored changes cause incidents or rollbacks
Rework rate	How often finished work has to be redone
Human intervention rate	How often a gate or escalation was needed, and whether that falls over time

Take a baseline for each metric before the rollout. Adoption is irrelevant if it isn’t measurable.

The better question

Stop asking whether the model knows that 97 is prime. Ask what share of real tasks it completes inside your harness, at what cost, and how often a human has to step in.

Models will keep improving, and you do not control that curve. You control the harness. That is where the engineering lives now.