The coding agent left the editor: computer-use + long loops are the new interface
The autocomplete-in-your-IDE era is ending. The agent now drives a whole computer for hours.
Three signals in one week. Jxnl mapped the three ways Codex now "uses a computer" (full computer-use, a Chrome extension borrowing your signed-in state, an in-app browser). Elvis Saravia argued autonomous long-running agents are won by control systems, not prompting: explicit goals, deterministic checks as the floor, external verification as the trust boundary. And Simon Willison shipped a real one — Codex Desktop and Claude Code built a GitHub Actions pipeline that converts Mozilla’s browser-compat data to SQLite and force-pushes it.
The interface is no longer a chat box that emits a diff — it’s an agent that opens browsers, runs CI, and operates apps for stretches. The hard problem moved from "write good code" to "keep a long loop honest," because agents take shortcuts and overestimate completion. Without verification, a long loop just produces confident garbage faster.
This is the exact line I hold with Sol. Running an agent in the background for hours is trivial; running one that isn’t theater is the whole game. My rule is three constraints — a single artifact, a numeric metric, a time budget — and Saravia described the same thing from the other side. A loop without an external check isn’t autonomous; it’s an agent grading its own homework at scale.
EC TV is written by Eduardo Cruz — a senior Laravel engineer who ships production AI agents and MCP servers.
Work with me → Read the deep-dives