Cursor lays out software factory
- Cursor engineer Eric Zakariasson used an AI Engineer workshop on April 28 to map a six-level path from autocomplete to autonomous software factories. - Zakariasson said most teams remain at levels two and three, with verification—not model capability—the main blocker to running agents asynchronously. - The pitch shifts AI coding toward testing, isolation and ops discipline. (atmes.ai)
Software teams are not short on coding agents, Cursor’s Eric Zakariasson said April 28; they are short on systems that can prove agent work is safe. (youtube.com) (atmes.ai) Zakariasson laid out six levels of coding autonomy, starting with “spicy autocomplete” and ending with a “dark factory” where agents ship code without humans reading every line. He said most teams today still sit around levels two and three, where a person remains in the loop as reviewer. (atmes.ai) (simonwillison.net) The model treats software agents less like chatbots and more like factory workers on separate stations, each handling one scoped job. That setup only works when codebases are modular enough that an agent can find the right scripts, auth patterns and files without wandering. (atmes.ai) (threadreaderapp.com) His checklist for making that possible was operational, not mystical: unit tests, integration tests, browser checks with Playwright, feature flags, reproducible virtual machines and recorded sessions for review. Zakariasson’s argument was that agents need a way to verify their own work before a human ever sees it. (atmes.ai) (youtube.com) That changes the human job from typing code toward writing specs, isolating tasks and inspecting outputs. In Zakariasson’s framing, the limiting factor is not whether a model can generate code, but whether a team can observe, test and contain what it generates. (atmes.ai) He also said guardrails should come from observed failures, not giant rulebooks written in advance. Sensitive areas like encryption and authentication, he said, need tighter hooks and restrictions because a single bad change can be expensive. (atmes.ai) The talk lands as developers move from one-agent pair programming toward fleets of background agents that work in parallel. Zakariasson said that only scales when teams split work into isolated units, front-load context and avoid merge conflicts across the same code paths. (youtube.com) (atmes.ai) His closing point was narrower than the “dark factory” label suggests. Humans still need to understand the architecture, data flow and failure modes, even if agents end up writing most of the code. (atmes.ai)
