Maintenance, handed to an agent
An agentic workflow does the maintenance, end to end. 100% agent work, 100% human supervision: it does, we approve.
The Problem
Maintenance is the work that never trends. Crash fixes, flaky edges, the slow decay a shipping desktop app collects. It is essential, easy to defer, and deferring it is exactly how a product turns fragile.
What We Built
We built an agentic maintenance workflow: the agent does the work, the team supervises. It runs on a small set of commands we drive, /goal to point it at what matters and /loop to keep it iterating, wired through agentic workflows that read production signal, choose the work, and open reviewable changes. /goal sets the target for the week, from Sentry and Amplitude signal to a specific reliability or performance goal; /loop keeps the agent working the problem, proposing changes, running the checks, and handing back only what passes. Every week it delivers a batch of finished work that we review, merge, and ship. Nothing merges without a human reading it, so the agent moves fast and we stay accountable. Performance climbed, the app got steadier, and customer feedback stopped piling up, because it gets triaged and closed inside the same loop.