AI Didn’t Get Smarter at Your Job. Someone Wrote Your Job Down.
Harry Glorikian
Author, The Invisible Interface (June 2026) | GP, Scientia Ventures | Research Affiliate, MIT Media Lab
May 1, 2026
I was at a friend’s 60th birthday a few weeks ago. Got to talking to his daughter. She works at a consulting firm. Her job is helping clients move into the AI world and think through strategy.
We ended up on the topic of harnesses.
You may ask, what is a harness.
A harness is the wrapper that goes around an AI model. It’s the rules. The tool list. The memory. The guardrails. The model itself sits in the middle, doing the thinking. The harness tells it which task to run, what data it can touch, when to stop, when to hand off to a human. Without a harness, a chat model is a smart generalist talking into the void. With the right harness, it does work that used to take a person hours.
Her clients aren’t asking for AI strategy in the abstract. They’re asking her firm to write the harness for one specific workflow. Customer support escalations. Insurance prior-auth. Lease review. Quarterly reporting. One workflow at a time. Each engagement a few weeks long. Each delivery a small piece of the company’s daily operations that now runs without the person who used to run it.
That’s a real job, today, paid for by real Fortune 500 budgets.
I’m telling you this because the AI conversation has gotten stuck on the wrong thing. Everyone is watching the next model. The next benchmark. The next demo. The thing that’s actually changing your industry isn’t necessarily another model release. It’s that thousands of people are quietly building harnesses around the models that are available today. I am trying to build one myself for one particular workflow I would love to automate around research.
The number to remember
A research nonprofit called METR measures how long a software task an AI can finish on its own. Their benchmark is software engineering and research work, not knowledge work in general. On that benchmark, GPT-5 (measured August 2025) handled about two hours and seventeen minutes of expert work, half the time, with no help. On a third of those tasks, it succeeded every single time. The doubling cadence on this metric was every seven months over the long run, and every four months in 2024 and 2025.
Software is the leading edge. The same trend hasn’t been measured cleanly in legal review, financial analysis, or clinical workflow yet, partly because the benchmarks don’t exist and partly because nobody has the budget for the human baselines. But a leading edge tells you something. Two hours of expert work, autonomous, on the kind of task most knowledge workers would still describe as “above my pay grade for an AI.”
Look at your week.
How much of it is two-hour blocks? The first draft of a memo. Reviewing a contract for standard issues. Building a deck. Pulling numbers and writing a recommendation. Triaging your inbox. The week is built out of two-hour blocks with meetings between them.
The model is already capable of doing a meaningful chunk of those blocks. The reason it isn’t doing them in your company is that nobody has built the harness yet.
The O-Ring problem decides who wins
I wrote a piece recently called “The Question I Keep Circling.” The argument, in one sentence: jobs are not piles of separable tasks, they’re chains of dependent ones, and Michael Kremer won a Nobel for showing how that changes everything. One weak link breaks the rocket.
When you put a harness on a job, you’re choosing which links in the chain get peeled off and run by the system. That choice is everything.
If the chain is short and the links are loose, the harness eats the chain. The work becomes a workflow. People get reassigned. Headcount drops. This is what happened to the CEO I talked to last month who had quietly automated his PR function and his website updates. Same titles on the org chart. Different reality underneath.
If the chain is long and the links are tight, the harness peels off the easy links and the human gets denser, harder, more leveraged work. Same job title. Different shape of day. Radiologists are at 81% AI use and their compensation went up nine percent last year. Customer service reps are at high AI use and the Bureau of Labor Statistics expects the role to shrink five percent by 2034. Same exposure number. Opposite outcomes.
The harness decides which one happens. The dimensionality of the job decides what’s left after.
Why this is happening now
Three things shipped in the last twelve months that made the harness possible.
Anthropic released a feature called Skills in October. A skill is a folder with a markdown file inside. The file says, in plain English, here’s how this task gets done. The model reads it when the request matches. It doesn’t memorize anything. Next request, different file.
A year earlier, Anthropic released the Model Context Protocol. MCP is the wiring that lets a model touch your calendar, your email, your CRM, your file system, your EHR, without a custom integration for every model and every vendor. OpenAI adopted it. Google adopted it. Microsoft adopted it. Over sixteen thousand of these connectors exist publicly today.
The harness is what wraps around all of it. Stripe has one for code review. Shopify has one. Airbnb built one specifically to migrate a giant chunk of their JavaScript to TypeScript. That migration would have taken hundreds of engineer-hours done the old way. They ran it as a workflow.
This is the part the headlines miss. The model didn’t get smarter at your job. Someone wrote your job down. The wrapper got good. And the wrapper is what eats the chain.
What replaces what gets eaten
The easy story is that AI eats jobs. The other easy story is that AI creates more jobs than it destroys, so relax. Both miss what’s actually happening.
The World Economic Forum’s 2025 Future of Jobs report surveyed employers covering fourteen million workers. The headline numbers: 170 million new jobs created by 2030, 92 million displaced. Net positive 78 million. Inside that net number, 39% of workers’ core skills change by 2030. PwC analyzed close to a billion job ads. AI-skilled workers in the same role earn a 56% wage premium, more than double last year. Skill change is fastest in the most automatable jobs.
The new categories are real. AI engineer postings up 143% year over year. Prompt engineer up 136%. AI content creator up 135%. Add the categories the Bloomberg piece on Mercor, the company that pays experts hourly to teach AI models how to do their jobs, exposed: AI trainer, harness builder, model auditor, evaluator. My friend’s daughter has one of those jobs. Mercor has thirty thousand contractors doing another version of it.
So new jobs do get created. The hypothesis holds. But the aggregate hides the distribution, and the distribution is where the real story is.
The O-Ring logic doesn’t go away when new jobs appear. It moves into the new jobs. The harness builder writes the chain that eats someone else’s chain. The AI trainer teaches the model the workflow they used to run themselves. The value of the new role depends on the same complementarity Kremer described. Some of the new jobs are high-leverage and well-compensated. Others are gig-economy task work paid hourly with no continuity and no path. The Bloomberg piece showed both kinds inside the same company.
The radiologist whose comp went up nine percent and the displaced customer service rep both sit inside the WEF’s net-positive 78 million. That’s the trick. If you’re inside an organization, the question that matters is not whether your industry nets out positive on a global spreadsheet. It’s whether the harness being built around your role is concentrating your work or eroding it, and whether the institution you’re in has a transition pathway or just a severance policy.
The value capture is the other piece. The person who writes the harness keeps more of the value than the person whose work gets harnessed. My friend’s daughter sees this clearly. Her clients are starting to. The Mercor contractors who teach models for hourly pay see it least.
The model didn’t get smarter at your job. The wrapper got good. The wrapper is being written this quarter, by someone, for some workflow, with a choice about which links in the chain stay human and which don’t. That choice is where the next decade of knowledge work gets decided. Not whether jobs disappear. How they reshape, and who benefits.
Sources: METR, time-horizon tracker (metr.org/time-horizons); Anthropic, Agent Skills launch (October 2025) and Model Context Protocol (November 2024); public engineering posts from Stripe, Shopify, and Airbnb on internal coding-agent harnesses; Mitchell Hashimoto on the formalization of “agent harness” terminology (February 2026); Michael Kremer, “The O-Ring Theory of Economic Development,” Quarterly Journal of Economics (1993); World Economic Forum, Future of Jobs Report (January 2025); PwC, 2025 Global AI Jobs Barometer; Autodesk, 2025 AI Jobs Report; Bloomberg Big Take, “The startup turning your job into AI training data” (April 2026); Bureau of Labor Statistics occupational projections; Medscape Radiologist Compensation Report (2025).