Cognition launches Devin Oncall: 100% on SREGym and a production fix in 2m 21s

Cognition has opened general access to Devin Oncall, a cloud on-call agent that triages alerts and bug reports on its own, diagnoses production failures, and attempts to fix them. According to the company, it scored 100% end-to-end success on the SREGym benchmark, with a best time to resolution of 2 minutes 21 seconds. In production use, the system already processes thousands of alerts a day for large customers.

The benchmark: what exactly was measured

SREGym tests agents on typical faults in live systems. Each test is based on a real incident, and in a single run the agent has to both find the root cause and apply a working mitigation. According to Cognition's charts, Devin Oncall reached 100% on GPT-6 Astra and Claude Opus 5.5, scored 94.1% on GPT-6.1 Sol, and 85.3% on Claude Fable 5.1. For comparison: Claude Code on Sonnet 5 scored 45.1%, Codex on GPT-5.6 Sol scored 76.5%, and CloudThinker on Opus 5 also scored 76.5%.

An important caveat. Cognition itself admits that mid-tier models have already hit the ceiling on this cleaner test set, while the previous generation scored 50–60%. The benchmark is saturated, and 100% on it does not guarantee the same result in a tangled legacy codebase. The company says outright that agents still tend to stop at the first plausible cause.

How the agent works

The design philosophy follows the "bitter lesson": instead of long, hand-tuned prompts, graph architectures, and fine-tuned SRE models, the bet is on a strong coding agent with the right context and tools. The authors, who reportedly include TierZero alumni, say that older tricks stop working as models improve and have to be stripped out. Devin's architecture is not tied to a single model, so the best one can be picked for each on-call task.

A second agent, a reviewer, guards against premature conclusions. It asynchronously checks the main agent's reasoning and tool calls and catches typical mistakes: correlation mistaken for causation, ignoring contradictory telemetry, and confusing the trigger with the root cause. By the system's logic, a fresh deploy is a lead, not an explanation.

Context is built from three layers:

• Per-Responder Memory: each responder's memory, for example whom to tag and how to behave in future investigations.
• Per-Org Skills: skills that Devin generates by regularly indexing the environment. Example: calculating the blast radius for customers.
• Per-Org Memory: post-processing of every investigation and a "low-fidelity cache" of institutional knowledge.

There are two products. Responders watch a stream of alerts or bug reports. A hidden parent agent decides whether the problem is new and, if so, launches a child investigation with fresh context. This skips about 60% of investigations (in the illustration, 65 out of 100 alerts), and costs drop to 46% of the "investigate everything" mode, a saving of 54%. Incidents joins the incident channel and maintains a live document with a timeline and evidence. It hands engineers' direct questions to child agents while the parent continues the background investigation. Messages are pre-filtered.

Customers and numbers

• Reducto: Devin spotted a trend across several alerts and opened an incident on its own, before the production alerts fired. It now triages most alerts first, sometimes closes medium-severity issues, or prepares and tests a PR before an engineer is paged.
• RV Tech (a Rivian–Volkswagen joint venture): the agent extracts vehicle IDs and timestamps from support requests, pulls logs, runs parsers, and does triage. One team moved several engineers back to product work.
• AngelList: it finds the root cause 3.5 times more often than the internal system Sky; 50% of incidents are resolved fully autonomously. Sky was retired.

Keep in mind that all figures come from the vendor's blog and have not been independently verified.

Practical takeaway for teams: it makes more sense to start with Responders on a single noisy alert channel and measure the share of skipped duplicates and root-cause accuracy. Connecting observability (reviews mention Datadog MCP) and code tracing will shape diagnostic quality more than the choice of model. Permissions for actions in production should be limited to mitigations such as rollback and scaling, as Cognition itself recommends, and any code changes should go through human review.