What it takes to run coding agents in production across your team.
Every team has a coding agent that ships great code on someone's laptop, and that stops working once the whole team relies on it. Drawing on what we learned running (and building) a software factory at Upsun, we'll dig into a few problems: sandboxing untrusted code, keeping non-deterministic models on a deterministic track, giving agents shared context instead of per-developer prompt files, and evaluating agent work beyond "the tests passed.