• The design choice that does the work is a refusal: River will not answer direct messages. She declines and suggests opening a public channel, which is what makes every request searchable, joinable, and reviewable by anyone in the company.
• The usual worry about AI is that the agent does the thinking. Lütke reframes it: the danger is not the agent doing the work, it is the work happening in a private window, where the person at the keyboard is the only one who learns anything.
• The merge rate climbed without a retrain or a model swap. It moved because people watched the agent work and fixed the patterns — a system improvement, not a model improvement.
• Context is pushed down to the edges. Each channel pre-loads the zones, skills and instructions its own team needs, written by the people closest to that work, and the agent keeps a memory it updates and un-learns over time.

• AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
• context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

River shows what happens when you design an AI coding agent around organizational transparency rather than individual productivity: every River conversation is searchable, joinable, and becomes institutional knowledge rather than a private exchange that disappears. If you're thinking about how to deploy AI agents inside a company without creating knowledge silos, this is the clearest real-world model for the alternative.

• articlePushing software engineering limits with “napkin math”Gergely Orosz
• podcastThe rise of the professional vibe coder (a new AI-era job) | Lazar Jovanovic (Professional Vibe Coder)Lenny Rachitsky
• videoPrototyping as Leadership: How a CTO Ships with AI Agents — Hursh Agrawal, The Browser CompanyAI Engineer

Checking sign-in…

Loading comments…