17% use OpenAI's agents — the harness is the real gap

The most useful number in AI right now isn't a benchmark. It's this: 98% of OpenAI's own employees used Codex in June, while only 17% of its organizational subscribers and under 1% of its individual subscribers touched the agentic tool at all. Same model. Same price. Two completely different outcomes. That gap is the whole story of AI agent adoption in 2026, and it is not a model quality problem.

What actually happened

TechCrunch reported the figures on August 24 in a piece on OpenAI's agent push. OpenAI released ChatGPT Work last month — a modified version of Codex aimed at non-engineers, priced at $20/month on the company's lowest paid tier, meant to run multistep work end to end instead of answering questions.

The distribution math is brutal. The joint app counts about 20 million users. ChatGPT itself has over a billion people typing prompts into it. Agents are a rounding error inside the most widely distributed AI product on earth.

OpenAI's engineers are candid about why. Akshay Nathan, a product engineering lead, put the constraint on the model side: "We're actually quite limited by our ability to parse everything that's available to us, and then take action on it." Joe Gershenson, who works on the harness, described the fix as scoping: "The goal of good harness engineering is to be more precise about what information the model really needs to solve your problem."

Why the agent adoption gap matters for your business

The 98% number is the tell. OpenAI's staff aren't using a better model than you can buy. They're working inside a codebase, ticket system, and toolchain the agent is already wired into, with people who know what to hand it. The harness — the tools, the data access, the boundaries — is the product. Everything else is a subscription.

Seats are not adoption. If you buy twenty $20 seats and 17% of them ever run a task, you did not deploy an agent, you bought a nicer chat window. Price the workflow before you price the license. Pick one repeatable, high-volume task — order exceptions, quote follow-ups, invoice chasing — and instrument it.

Scope beats capability. Gershenson's line is the entire job: decide precisely what the model needs to see. Every agent we ship that works, works because it has three tools and one job, not forty tools and a mandate.

Source: TechCrunch.