Microsoft's ArgusAgent Keeps Coding Agents Running for Weeks

Most coding agents run in one straight line: prompt, code, done. Microsoft's ArgusAgent is built to do the opposite, staying on a task long after a single answer would have stopped. That is a different kind of tool, and the numbers Microsoft is reporting for it are worth a look.

The shift is in the loop itself: Plan, Execute, Verify, Review, Continue. Instead of finishing one task, Argus is meant to keep an entire research or engineering campaign running for the long haul.

What Happened

ArgusAgent splits the work across four roles that hand off to each other: Manager, Planner, Engineer and Reviewer. A run moves through planning, execution, verification and review, then loops back and keeps going.

Microsoft says the agent is aimed at long-horizon research and engineering, and that it reviews its own output independently for correctness and evidence instead of trusting the first result. It can pause, resume and recover, and it keeps persistent skills, memory and evidence across a campaign.

It works with Codex CLI, Claude Code, Copilot CLI, Cursor CLI, OpenCode and more, and it gives you both a web UI and a terminal cockpit.

The reported numbers: 1,548 hours across 27 campaigns, with a human research decision needed only about once every 310 hours and a 95–99% duty cycle.

Why It Matters

That duty cycle is the part to watch. If those results hold up, the pitch stops being "AI that writes code" and starts being "AI that runs the engineering loop" — something you leave running and check on, rather than something you babysit prompt by prompt. It is an open-source project, so the code is out there to read.