Most of us still "use" AI. Open a chat, type a prompt, get an answer, close the tab. The author argues that era is ending. AI agents have become productive enough that getting work out of them is no longer the hard part. The hard part is finding enough human attention to check what they have done. Once you see that clearly, you stop treating AI as a tool you pick up and start treating it as someone you hire.
The lesson starts with giving every agent its own computer. One team tried running five agents in the same code checkout, and the results were chaotic: one agent stashed everyone else's work, and another wiped the checkout entirely. Their fix was to give each agent its own virtual desktop, an isolated container running a full Linux desktop with its own file system, browser, terminal and code editor. Agents can see what they build, so front-end work can be tested in a real browser. Because the desktops live on shared servers, a developer in Tokyo can hand off an agent to a developer in London exactly where it left off.
The team began by building an on-premise alternative to OpenAI, but customers kept asking what they should actually do with it. The answer was mostly coding, and AI coding tools were not built for teams. So they built a kanban board where agents do the work. Each card is a task about the size of a user story. Agents read the codebase and write a spec, a human reviews it, the agent builds what was approved, and the work passes code review before merging.
The approach has limits, and the author is candid that it can fall apart. Still, the core shift is clear: the bottleneck is human review, so the job becomes managing and checking work rather than prompting for answers.
Source: Towards Data Science · Summarized by HeadlinesBriefing