Coding went mainstream. The results didn’t. Welcome to the age of the bot sitter.
Half the world writes code now. The P&L mostly hasn’t noticed. Because the real job quietly changed — from writing the work to supervising the machines that write it.
he adoption war is over. Roughly nine in ten US developers use AI coding tools daily; GitHub puts AI’s share of new code at about 46%; nearly two-thirds of vibe-coding users aren’t developers at all. Building software stopped being a specialist act sometime last year — a founder, a marketer, an operations manager can now ship a working tool in an afternoon. Coding, in the loosest sense, went mainstream for the world.
And yet. Forrester finds only 15% of AI decision-makers can point to an EBITDA lift. MIT’s 95% figure for pilots without P&L impact still stands. Inside engineering teams the fine print is stranger: code churn is up 41%, duplication has quadrupled, 63% of developers say they’ve spent longer debugging AI-generated code than writing it themselves would have taken — and trust in AI output has fallen while usage climbed. Everyone is producing more. Fewer can prove it amounts to more.
The job that appeared while nobody was hiring for it
Look at where the tokens actually go and you can see the new job forming. In multi-agent development pipelines, the review stage — not generation — consumes the largest share of compute; most of the budget is spent re-reading context and checking work, not producing it. The same is true of the humans. The developer who used to write functions now specifies tasks, watches three agent sessions run, reads diffs, rejects one, re-prompts another, approves the third. Karpathy — who accidentally named the vibe-coding era — has already moved the goalposts to “agentic engineering”: humans set architecture and judge output; machines do the typing.
We call it what it is: bot-sitting. It isn’t a lesser job — supervision is precisely the skill the 95% of failed pilots lacked. But almost nobody has staffed for it. Companies bought the agents, kept the org chart, and are quietly surprised that output stalls at the one person who can competently review what twelve agents produce. The bottleneck moved from typing speed to judgement — and judgement doesn’t scale by subscription tier.
The mainstream-coding era won’t be won by whoever runs the most bots. It’ll be won by whoever supervises them best per hour of human attention. If your agents are multiplying faster than your ability to check them, that’s a delivery problem — and delivery problems are the ones we fix.
If you run a team of bot sitters — knowingly or not
Staff the supervision, not just the subscription. Every agent you deploy creates review work somewhere. If nobody owns it, it lands on your best person’s evenings — that’s the stall you’ll feel three months later.
Design for review. Small diffs, tests as gates, one metric per task. An agent whose work can be checked in minutes is worth ten whose work takes an afternoon to distrust.
Measure outcomes, not activity. Prompts sent and code generated are vanity metrics. Tickets closed, cycle time, defects shipped — the boring numbers are where the 15% who see EBITDA lift live.
Train the sitters. Reviewing machine output well is a teachable skill — specifying, sampling, spotting confident nonsense. Our training practice coaches exactly this, one person at a time.