The Work Didn't Get Smaller. The Worker Just Got Faster.

Two Stories, One Workplace

If you only read the macro forecasts, you would think the AI era is a clean productivity story. A new working paper from Brookings and the Federal Reserve, reported by Fortune on July 2, projects that an AI productivity surge could shave $2.2 trillion off the U.S. deficit by 2036. The same paper notes, almost in passing, that more than half of those gains would be clawed back by side effects: a smaller labor force, longer lives, higher borrowing costs, an AI arms race. Goldman Sachs, in a July 3 report titled "An AI Job Apocalypse?", runs the same logic one more turn. Big productivity, big displacement, and somewhere in the middle, a tax bill.

If you only read the engineering press, you would think the AI era is a clean labor story. A June 30 LeadDev report found that 45% of engineers are now working more hours per week than a year ago, up from 38% in 2025. The biggest jump is at the top of the ladder: 53% of staff and principal engineers report longer weeks, nearly double the 2025 number. Almost half of all software engineers say they feel emotionally drained at work at least once a week. CTO burnout is up 30 percentage points in a single year.

Same economy. Same quarter. Very different picture.

The Agent Is Doing Its Job. The Job Is Just Bigger.

The reconciliation is hiding in plain sight. AI is doing exactly what the productivity forecasts said it would: more output per worker, faster cycle times, higher throughput. The Brookings paper estimates that AI could add the equivalent of a full percentage point to annual GDP by 2036. The agent benchmarks are catching up to senior-engineer performance on a wider class of tasks every quarter. Zuckerberg told Reuters on July 2 that agent development is going slower than expected, but slower-than-expected inside Meta in 2026 is still a category of progress that did not exist in 2024. The technology is working.

What is not working is the human side of the contract.

The LeadDev report quotes Steve Yegge, who has been building with these tools longer than almost anyone, on what he calls the "AI vampire" effect. The pattern is consistent enough to have its own name. An engineer sits down to ship one feature. The agent generates the code. The engineer reviews, refines, prompts again, reviews again, ships the PR. The next ticket is already loaded. The session has no natural stopping point. There is always another problem with an immediate next step. The dopamine loop, with its small wins, occasional surprise wins, and steady stream of progress signals, keeps the engineer at the keyboard long after the work is done.

"The AIs can be like sirens, and can woo you into staying at your computer longer than you should."

A LinkedIn commenter in the same piece called it more bluntly: it is a slot machine. You put in tokens, hope for a jackpot, and when you do not get one, you put in more tokens. The agent is doing its job. The job is just bigger than the engineer, and the engineer cannot stop feeding it.

The Work Is the Same. The Bill Is Bigger.

The work itself is what has changed. Pre-2024, a senior engineer's day was about 60% doing and 40% coordinating. Today, with Claude Code or its cousins in the loop on every commit, the ratio has flipped. The doing is cheap. The coordinating is now everything: planning the agent's work, reviewing its output, recovering from the cases where it confabulated a library, re-prompting when the spec was vague, writing the spec so the next prompt is not vague, explaining to your manager why the PR took two hours of review and ten minutes of code.

Berkeley researchers have a name for this too. When productivity tools make "doing more" feel achievable, users take on more work, work faster, and multitask excessively. The agent is not replacing the engineer. It is replacing the engineer's excuses. Every wall that used to stop a session, including waiting on a review, hitting a hard problem, and running out of steam, is gone. The work has expanded to fill the time the agent freed up, and then some.

The Brookings paper sees this at the population level: a projected 3% drop in labor force participation, roughly 6 million fewer people working by 2036, on a scale similar to the COVID shock but permanent. The LeadDev report sees it at the individual level: a 30-point jump in CTO burnout, a near-doubling of senior engineers working longer weeks, an industry-wide slide into longer hours dressed up as productivity.

There is a version of the future of work where the agent handles the boring half, the engineer handles the interesting half, and the interesting half gets smaller every quarter. That future is the one on the cover of the productivity papers.

There is another version, the one most engineers I know are living in 2026, where the agent handles the boring half, the engineer handles the interesting half, the interesting half gets bigger, and the engineer also handles the management overhead of running the agent. The same engineer, the same laptop, the same calendar, the same salary. A work week that quietly grew by 8 to 12 hours. That is an absorption story, not a productivity story. The system got more capacity, and the system used that capacity to ask for more output.

The future of work, in other words, is not the future the macroeconomic models were modeling. It is a future where the work is the same work, the agent is the same agent, and the only thing that actually changed is who is on the hook for what the agent produced. The Brookings paper calls it a $2.2 trillion deficit fix. The engineer in the LeadDev report calls it Tuesday.

The agent is doing its job. We just have to do ours.