Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Claude Code Head Boris Cherny: My AI Booked Eight Flights And Five Hotels Autonomously
TL;DR
Claude Code plus Claude Co-Work handled a real multi-city travel job end-to-end — Boris Cherny says the agent checked his email and calendar, caught two missing stops and wrong dates, then booked eight flights and five hotels in about an hour, with only one hotel needing a rebook.
The biggest mindset shift is that AI capabilities are changing month by month, not year by year — Cherny argues engineers who last tried coding models a year ago are judging a different product entirely, and says users need to keep a “beginner mindset” because the next model may suddenly nail tasks that previously failed.
Anthropic says Claude Code drove a 250% increase in code written per engineer without hurting quality — Cherny contrasts that with the old world at Facebook/Meta, where a 1-3% annual productivity gain was considered a major win and took huge effort to achieve.
Cherny dismisses “token maxing” as a major driver of demand, even as Alex Kantrowitz cites Amazon abuse cases — Kantrowitz quotes a Financial Times report and an Amazon employee who said they ran pointless automations for hours just to hit AI usage targets, but Cherny says Claude Code usage is broad-based across many customers, not concentrated in one account.
His advice to companies is simple: give people tokens, give them psychological safety, and don’t optimize too early — Cherny says the best AI wins often come from unexpected people, like an accountant, marketer, or new grad, not the engineers you’d have picked in advance.
The deeper thesis is that AI adoption looks like the computer adoption paradox all over again — Cherny invokes a Harvard Business Review argument from the 1990s: companies only saw real gains from PCs once they rebuilt workflows around them, and he thinks AI will require the same painful organizational redesign.
The Breakdown
Eight Flights, Five Hotels, and One Wrong Neighborhood
Cherny opens with the kind of story that makes the whole “AI agent” pitch feel less abstract: he used Claude Co-Work to plan a five-stop trip around events in London and Tokyo. The system checked his email and calendar, found two stops he had missed and corrected dates he had given wrong, then booked eight flights and five hotels while he went back to coding. One hotel landed in the wrong area, he asked for a rebook, and that was that.
Why Old Mental Models About AI Are Already Stale
He says this was the best result he’s ever gotten from one of his recurring test cases, and he credits the combo of Co-Work with Opus 4.7. The bigger point is psychological: people who formed an opinion of coding models a year ago are often badly out of date, because this is the first technology he’s used where capability seems to jump every month. His advice is basically to keep re-testing assumptions, because the thing that failed before may now work “perfectly right.”
From Software Interfaces to Agents That Work Around You
Alex Kantrowitz reframes the shift as moving from fixed software interfaces to something more personal and adaptive. Instead of clicking through bloated websites built for scale, the user can now delegate to an agent that knows preferences and goes out to do the work. Cherny agrees this is the core of why people are responding so strongly and why growth looks explosive.
The Token Maxing Challenge Gets Put on the Table
Kantrowitz then pushes on the hype by raising “token maxing,” where companies reward employees for burning lots of AI usage. He describes leaderboards, quotas, and a broader fear that demand may be inflated by bad incentives rather than real utility. Cherny’s immediate answer is that he doesn’t think token maxing accounts for a large share of usage.
Why a 250% Productivity Jump Feels So Radical
To explain why demand feels real, Cherny reaches back to his time at Facebook, where he worked on code health across Facebook, Instagram, and WhatsApp. In that world, a 1-3% productivity improvement per engineer over a year was a hard-won success; with Claude, he says Anthropic has seen code written per engineer rise by roughly 250% while code quality and reliability stayed stable. That, in his telling, is why companies are scrambling to figure out how to capture similar gains.
His Playbook: More Access, More Safety, Fewer Assumptions
Cherny says companies shouldn’t force every token request through approval systems, and they probably shouldn’t over-focus on competitive token-maxing schemes either. What matters more is letting people experiment freely and creating psychological safety so they can change their own workflows without fear of looking foolish when something doesn’t work. He also stresses that the breakout wins often come from surprising places — an accountant, a marketer, a new grad — not the obvious “top performers.”
Amazon Abuse Stories vs. Cherny’s Broader Adoption Thesis
Kantrowitz comes back with receipts: a Financial Times report on Amazon staff using AI for unnecessary tasks to inflate usage, plus a direct quote from an employee who said they run pointless automations for hours and delete the results just to meet targets. Cherny doesn’t dispute that this may happen, but says Claude Code’s customer base is broad enough that no single company appears to be distorting demand. His answer is less “that never happens” than “that’s not the main story.”
The 1990s Computer Paradox, Replayed With AI
Cherny ends on a historical analogy he clearly loves: the old question of why computers were everywhere but productivity gains were hard to find. His summary of the answer is that companies only benefited when they rebuilt processes around computers instead of keeping them at the edges of paper-based workflows. He thinks AI is in that same phase now — the winners will be the organizations willing to redesign how work gets done, not just bolt a model onto old habits.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
