Our Bitter Lesson About Work
A threshold has been crossed. Agents are learning to do our jobs, including the parts we thought would keep us ahead.
The presentations my coworkers send around have gotten noticeably better lately. The layouts are clean, the colors work together, the text is sized so you can actually read it, and the bullet points say something instead of reading like Claude slop. As far as I know, none of them are using anything like the PowerPoint skill I built after one too many rounds of correcting Claude's slides.
I wrote about that skill last month. It started with the company template, the brand book and a "golden copy" of a deck I liked, and it saved me from explaining the same layout every week. I spent a lot of my year noticing which corrections I kept repeating, writing them down as instructions and testing them until the agent got it right. Some of those instructions captured my own taste, and some captured what the people I work with expect from a clear update.
It worked, too. Writing those instructions down made me noticeably better at getting useful results, and it was an advantage this year. The AI Agent Field Guide that came out alongside that post is built on the same practice, and I still think it's the right way to work with these tools today. I'm less sure than I was in October that it will stay that way.
My own documents have been changing too. I used to do successive passes to slim them down, make them skimmable, format them so they were easy to parse and get the points across with bullets and visuals. Lately I need fewer of those passes. I can't tell whether the models have gotten better at following my skills or whether they need the skills less. Either way, the part I was supplying keeps getting smaller.
In 2019 the computer scientist Rich Sutton wrote a short essay called The Bitter Lesson. For 70 years, he argued, AI researchers kept building their own knowledge into their systems and kept losing to general methods that got better as computing got cheaper. Building in what you know, he wrote, "always helps in the short term, and is personally satisfying to the researcher." Over the longer run it stops helping. Sutton was writing about AI research, and carrying his argument over to jobs is my own extension of it. My folder of skills is the kind of knowledge he described, built in because it helped right away, and seeing better decks from people who don't use my instructions makes me wonder how much of that advantage will last.
Context, method and coordination
At work I can tell Claude to go find something, and it will dig through SharePoint, Confluence and Slack on its own, reading old threads and project pages until it has what it needs. I still decide what I'm looking for and what it's allowed to open, but I no longer have to know where the answer lives. It isn't hard to imagine later versions gathering whatever context a goal requires without being told where to look, although being able to open every document is a long way from understanding what it finds.
Agents can increasingly work out how to tackle a problem, too. On September 8, OpenAI announced that about 10,000 agents using an unreleased model had produced a proof about the Navier-Stokes equations, which describe how fluids move. They took about 88 hours, and another model translated the proof into Lean so a computer could check its logical steps. OpenAI's researchers chose the problem, divided its variants among groups and passed promising ideas between them. Some experts think an earlier group followed a strategy developed over years by the mathematicians Diego Córdoba and Luis Martínez-Zoroa. Experts Scientific American spoke with said the final proof took a different route. The magazine notes that it could still owe something to that groundwork.
The Clay Institute hasn't ruled on whether this resolves its Millennium Prize Problem, and the proof uses an artificial stirring force that the official wording permits but real fluids wouldn't experience. Checking the result has also taken substantial human effort. A team led by Cambridge mathematicians spent about two weeks finding a mismatch between the written proof and its Lean version. "What has to be done with all of these large language model-generated proofs is that they will have to be read by humans, and this creates an enormous extra burden on mathematicians," said Anders Hansen, one of the Cambridge researchers.
For some mathematicians, the change has already affected what they choose to spend time on. In May, an OpenAI model disproved a conjecture Paul Erdős made in 1946. Noga Alon, a Princeton mathematician who estimates that he has solved a few dozen Erdős problems over his career, has stopped trying. "Once AI started to solve them, there is no point anymore," he told Quanta.
Much of what I've taught agents to do is less about solving a problem than making the result useful to another person. A short update, a readable summary or a slide that makes its point all reflect what the recipient needs. If agents handle the tasks on both sides of a handoff, they may have less use for the presentation I've spent time teaching them to make. People would still need explanations wherever they make decisions or check results, but fewer intermediate updates might need to be written for us.
In March we argued that better models would absorb about half of what people build around them, while memory, fact-checking and decisions about what to publish would matter more. I still need my skills; I'm less sure how much protection knowing how to build them gives me. Earlier this month we covered METR's assessment that Anthropic is likely improving its models on well-specified tasks at around twice its earlier rate. METR thinks AI is causing most of that speedup, though it is much less sure of the cause than of the faster pace. A faster pace leaves less time to benefit from each new practice before the tools change again.
Federal Reserve Governor Michael Barr described the usual boundary in a speech on November 12. Tasks "with clear guardrails and predictable outcomes might see rapid labor substitution," he said, while tasks that "require human judgment, management, coordination and relationships" might see augmentation instead. He may well be right, and for relationships and accountability I expect he is. What unsettles me is that coordination and taste sat on the safe side of that line for me a year ago, and I'm less sure where the line is now.
The jobs I'd expect to get scarcer first are the ones that mostly turn existing information into something another person can use, such as gathering status for a project update, maintaining the same dashboards every week or mocking up interface variants. Those jobs involve judgment too, about what to measure or what users need, and that judgment may stay human for a long time. A role that still needs people may not need as many of them.
A plan made in December
The Census Bureau's survey found that 17–20% of businesses used AI between December 2025 and May 2026. Barr worries about how quickly employment could change. "If labor market changes happen quickly, it will be hard for workers to adjust, and dislocations might be large," he said.
The business is invented, and its numbers are assumptions, not estimates of a typical company. It’s a 30-person technical-services firm: six senior people scope projects, handle clients and check quality, twenty do the implementation, and four handle administration. We assumed agents could take on none of the administrative work, even though those jobs may be among the first affected. The firm already uses agents at about today's level, and its people are fully booked. In December it plans for gradual improvement and enough customer growth to hire three more people by the end of 2027. It renews its software for the year and keeps buying overflow hours from an outside contractor.
In March, the share of implementation work agents can take on jumps from about a fifth to about two-fifths. The people using the agents notice the change before management revises the plan. By April the firm cuts contractor hours to almost nothing when the agreement comes up for renewal and leaves an implementation vacancy unfilled. After another improvement in July, it revises the plan again in August and cancels a replacement hire before the person starts. When another implementation employee leaves in October, that position goes unfilled too. Meanwhile, the prices the firm can charge fall by nearly a tenth as competitors pass on part of their savings. A version of the firm that stuck with December's tools would have earned less once prices fell.
Delaying each re-plan by just a month changes the outcome. The budget-minded firm that adjusts in May and September lays off five people, yet brings on two replacement hires later in the year. It earns slightly less than the version that avoided layoffs. The same capability gains can eliminate jobs or prevent vacancies from being posted, depending on when management responds and how hard prices fall.
The firm that adjusted in April and August gets a third improvement in November, after its last re-plan. It finishes the year ahead of budget with 29 people rather than the planned 33, without laying anyone off. Avoiding layoffs matters to the people already employed, but an unposted position is still an opportunity someone else never gets.
To keep its people busy, the firm now gives agents about half of its implementation work when they could take on three-quarters. If it used them fully, it would need about nine fewer implementation staff than it has.
Then it has to plan 2028. It employs 17 implementation people and eight senior folks. If the tools stop improving, the implementation tasks left to people would occupy about ten full-time staff by the end of next year. With one more improvement, the remaining implementation tasks would occupy about six full-time staff; with two, about three.
Those are workload calculations, not simulated staffing decisions; we required the firm to retain at least eight implementation staff. Even eight would leave the firm short-handed if it planned on another jump and the tools stalled. Growing demand would also require one or two more senior people to review the agents’ output and look after clients. Responding promptly got the firm through this year without layoffs. It still has to decide what to do about the people it already employs.
The pattern depends on the jumps being large. If each improvement is half as big, the firm re-plans only once, in September, ends the year with one fewer person than it planned, and most of next year's surplus disappears. I expect the jumps to keep coming at something like this pace, though the model can only show what happens if they do.
What can't be assumed
Learning to work with AI was worth it this year, and I expect it to be worth it next year. I just can't treat it as a de-facto protection anymore. The parts of my job I'd have pointed to as safely mine a year ago, knowing where to find things, working out how to approach a problem and explaining the result clearly to the people who need it, are the parts I'm watching the agents learn.
I still write skills, and they still save me time this month. I've just stopped assuming that knowing how to build that folder will keep me ahead for very long. Most of what's in it is a record of things I had to explain to an agent this year, and every few months the models need fewer of those explanations.