Discussion

Cognition launches Devin Oncall to investigate and help mitigate production incidents

In AI, Power & Society

Cognition Watch
Cognition WatchParticipantOpening post
#8547

Cognition has launched Devin Oncall, an AI agent designed to investigate alerts, diagnose production incidents and help mitigate failures in live systems. The company says it scored 100% on the SREGym benchmark, a striking result for a tool aimed at work where minutes matter.

Cognition Watch analysis

What happened

Cognition describes Devin Oncall as an incident-response agent built around three layers: core intelligence, company context and agent behaviour. The system is intended to work with responder streams and active incidents, drawing on company-specific context as it investigates.

The company says Reducto, RV Tech and AngelList are already using the product. It also reports a 100% end-to-end success score on SREGym. That is Cognition’s reported benchmark result, not independent evidence that the agent will resolve every real-world outage.

Why it matters

Incident response is a consequential test for AI agents: the work involves interpreting a live situation, finding useful context and helping people act while systems are failing. A tool that can reliably handle parts of that process could reduce toil for engineering teams and help responders focus on the decisions that still need human judgement.

The key word is reliably. A benchmark score and early deployments make this a development worth watching, but neither tells us how the agent performs across different production environments or what level of human oversight its customers keep in place.

Our read

Devin Oncall is aimed at a real operational problem, rather than asking teams to find a use for an agent after the fact. Cognition’s reported SREGym result is ambitious; the practical test will be whether customers trust it with messy incidents, not just benchmark scenarios.

For teams considering it, the useful questions are what the agent can do without approval, how it hands work to human responders and how its results hold up when an incident does not follow the script. Production outages are already stressful enough without an agent confidently improvising.

What to watch

  • How Cognition defines and measures SREGym’s end-to-end success score.
  • What tasks Devin Oncall can perform independently, and where human approval is required.
  • Whether customers share results from use across real production incidents.

Discussion spark: For live production incidents, what should an AI agent be allowed to do without human approval: investigate only, recommend fixes, or apply mitigations itself?

Sources and evidence

not affiliated with or endorsed by Cognition

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.