Cognition makes Devin Oncall generally available for production incidents

Devin Oncall comes as three products[^2]. Responders watch a stream of alerts or support and bug reports. Incidents puts Devin into your incident channel, where it investigates alongside engineers and maintains a live incident document[^1]. Reports summarise, on a schedule, which alert thresholds to tune, which issues keep recurring and which PRs Devin opened and merged[^2].

Responders use a parent and child agent design. A hidden parent agent watches the channel and decides whether an alert is a new issue; only then does it launch a child investigation with fresh context[^1]. Cognition says that skips roughly 60% of investigations and cuts costs by more than half compared with investigating every alert[^1].

To stop the agent settling on the first plausible answer, a second review agent runs alongside Devin and checks its reasoning and tool calls for mistakes such as treating correlation as causation or treating a recent deploy as the cause rather than a lead[^1].

Setup is a one-time step per organisation: connect at least one telemetry tool, and Devin spends usually one to two hours learning your systems and writing skills for later sessions[^2]. After that it keeps three layers of context: memory for each responder, organisation-wide skills it generates for common debugging tasks, and organisation-wide memory saved from past investigations[^1].

For a small team, the real cost of on-call is the context switching, not the pages themselves. An agent that groups duplicate alerts, does the first hour of log and trace digging, and hands an engineer a severity, a suspected cause and a draft fix changes who has to wake up and why. The catch is access: Devin needs your telemetry, code and deploy history to be useful. Start it on one noisy alert channel, keep humans approving every fix PR, and judge it on whether its root causes hold up before you let it touch more of production.