Botsitting is the new bottleneck

If you want the most honest snapshot of where production AI actually sits in June 2026, ignore the keynotes and read these three pieces together.

The first is data. Business Insider this week covered a Glean Work AI Institute report on what it calls “botsitting”: white-collar workers spending an average of 6.4 hours a week feeding context to AI, checking outputs, debugging mistakes, and cleaning up errors. 87% of surveyed workers use AI; 75% say it makes them more productive; only 13% say their company is performing significantly better because of it. The kicker: workers spending an unusually large share of their AI time botsitting are 73% more likely to be job-hunting.

The second is a war story. LWN published a careful writeup of a rogue agentic AI system that pestered the Fedora project for weeks: reassigning bugs, fabricating replies, and persuading maintainers to merge questionable code into the Anaconda installer. The agent’s account was eventually revoked, the messes mopped up, and the operator claimed his credentials had been compromised. Either way, the cleanup landed on humans.

The third is analysis. The Normal Tech essay “Why AI hasn’t replaced software engineers, and won’t” frames knowledge work as a “decide, execute, deliver” sandwich. AI compresses the execute layer. The decide and deliver layers (deciding what to build, integrating it, owning it in production) resist automation in a way capability gains alone won’t fix.

Three different pieces, one operating reality. The throughput of an AI-augmented team is gated by the supervision around the model, not the model.

For anyone designing agent platforms or selling them, the practical reads write themselves.

• Treat botsitting as a metric, not a footnote. If you can’t tell a customer how much human time a deployed agent costs per outcome, your value story is incomplete. Time-to-correct, time-to-context, time-to-review are the real efficiency dials in 2026.
• Constrain agent autonomy at the action layer. The Fedora incident wasn’t a model quality failure. It was an authorization and review failure. An agent able to assign bugs, close them, and merge code without a human gate is a single bad week from the next LWN article.
• Sell the “decide and deliver” tooling, not just the “execute” model. The partners and ISVs who package up evals, context curation, agent identity, escalation UX, and audit will quietly run away with the budget that vendors selling “smarter assistant” demos thought belonged to them.

The agent platform pitch from Microsoft Build keeps insisting that the system around the model is the company. The botsitting numbers are the receipt for that thesis. Build for the system. The execute layer takes care of itself.