Everyone uses the agent; few let go of the work

Two numbers, same developers, same report. They use AI in roughly 60% of their work. They’re comfortable fully handing a task over to it in 0 to 20% of cases. Almost everything interesting about AI and careers right now is happening in the space between those two numbers.

They come from Anthropic’s 2026 Agentic Coding Trends Report, out this summer. If you skipped the coding news: a coding agent is AI that takes a whole task and runs it across a real codebase, writing the code and testing it and fixing what breaks, while the developer supervises. Working alongside one has become ordinary. Anthropic reports that 86% of organizations have moved coding agents out of pilots and into production, and that 42% now trust an agent to lead development work with a human checking it. Giving an agent a task and then leaving it alone is a much rarer thing. That’s what the 0 to 20% measures.

The distance between using an agent and handing work to it is where career outcomes are now separating, and none of the instruments the labor market runs on can see it.

Look at what those instruments do measure. PwC’s 2026 Global AI Jobs Barometer puts the wage premium for AI-skilled roles at 62% over non-specialized peers. Inside that premium sits a sharp step: postings asking for two or more AI skills pay a 43% premium, while one AI skill pays 28%. The market prices depth of fluency at a threshold, and every comp band built on that data sorts people by how much AI skill they can demonstrate. Postings count declared skills. Salary surveys count titles. Both are measuring fluency, and fluency is the thing heading to 60% of everyone’s workday.

Once most people in a market can clear a bar, the bar stops telling you who’s who. Wage premiums will keep paying out for a while, because postings and surveys still ask the question that way. The sorting power goes first, and the price follows later.

What Anthropic finds sorting people instead is a behavior. Developers comfortable with full delegation on 15 to 20% of tasks report optimism, and describe their job as coordinating agents and judging the output rather than producing it themselves. Developers stuck at ambient use report anxiety and a sense that their options are narrowing. Anthropic reads this as two groups rather than one gradient, with a threshold in between that changes how people describe their own future. The same report puts a generational wedge on top: 40% of junior developers report a perceived probability of losing their job, against 10% across the surveyed population.

Hold that 40% next to the most-quoted reassurance in the field. Lenny’s Newsletter surveyed between 6,000 and 8,200 tech workers this year and found job loss named as a primary fear by only 22%, well behind pace and pay. That number gets read as evidence that the fear is overblown. Both numbers can be true at once. If juniors fear job loss at 40% and everyone else at 10%, the population average lands right around 22%. The calm number and the alarming one describe the same workforce, and the calm one is the one that got published.

The same survey found career optimism falling from 54.8% to 48.7% year over year, which reads as a field getting gloomier together. If Anthropic’s split is real, that reading is wrong in an important way. A population pulling apart into an optimistic cohort and an anxious one produces a falling average too. I can’t prove which is happening from the published data, because nobody has cut the optimism measure by delegation depth. That’s the point. The cut that would settle it is the cut nobody makes.

Now the part that should bother anyone advising juniors. Stanford’s SIEPR payroll analysis found 22 to 25 year olds in AI-exposed occupations down 16% in employment, while seniors in the same roles gained 6 to 9%. The usual explanation is that AI automates the task types that make up junior work. Set that against the delegation finding and something recursive appears. Delegating well takes knowing which work is safe to hand over and being able to tell whether what comes back is any good. That judgment is what seniority buys. So the behavior that tracks optimism is the one juniors are worst placed to adopt, which makes it a poor rescue for the group taking the damage. That’s my inference from two findings sitting side by side. Neither source claims it. But it predicts something checkable: internal AI training programs, which PwC credits with a 3 to 4 times lift in adoption, should raise usage among juniors without moving handover much. Usage is the part that no longer sorts anyone.

You should be suspicious of all of this, and the reason is sitting in plain sight. Anthropic sells the agent. Anthropic measures how deeply people hand work to the agent. Anthropic reports that the people who hand over the most feel best about their careers. That’s a vendor grading its own category, and the 42% trust figure is a vendor asking whether you trust its product class.

Discounting it on those grounds runs into an awkward problem. Delegation depth doesn’t show up in a posting or a payslip. It shows up in telemetry. Anthropic can measure it because it counts sessions: its June Economic Index work cross-checked 9,700 survey respondents against roughly 235,000 tracked Claude Code sessions, and found a median of about one human prompt per task through the agent versus around 13 rounds of back and forth in chat. Measured that precisely, the variable exists only inside vendor logs, and the parties holding those logs are the ones selling the tools. Waiting for independent confirmation assumes it’s on its way. It may be structurally hard to get.

Which leaves the question that actually decides what to tell people. Anthropic’s own June report declined to say whether delegating makes people optimistic or optimistic people delegate more, and named both possibilities without picking one. That ambiguity has sat open since July, and it isn’t academic. If delegation builds confidence, then teaching people what to hand over is a real intervention, and the sharpest one available for the cohort down 16%. If confident people simply delegate more, then telling anxious juniors to delegate harder is handing them the symptom as the cure. The two readings prescribe opposite things to the same people.

The encouraging part is how cheap the missing evidence is. “What share of your tasks would you hand to an agent without checking the result?” is a question any survey could carry, alongside years of experience. Self-report is rougher than session logs, and it would still be the first reading of this variable from outside a vendor. As far as I can tell, nobody outside Anthropic is asking it, so the most important variable in this story sits unmeasured mostly because nobody has bothered. Your own team is a sample of one, and you could put the question to them this week. Ask for the number by tenure. One company-wide average will hold steady on the dashboard even while the two halves underneath it pull apart.