Cognition Releases Devin Oncall: An AI-Powered SRE Platform for Automated Incident Resolution
Cognition has launched Devin Oncall, an AI agent designed to autonomously investigate, triage, and remediate production incidents and system alerts.
Devin Oncall interface displaying automated incident investigation flows.
- Devin Oncall is a generally available AI agent designed to handle SRE-related production alerts and bug investigations.
- The tool integrates with existing monitoring stacks like Datadog, Sentry, and Dynatrace to perform automated root-cause analysis.
- Features include alert deduplication, automated log parsing, and the ability to open and test pull requests for incident mitigation.
Overview of Devin Oncall
Cognition has officially announced the general availability of Devin Oncall, an AI-powered agent designed to assist engineering teams in managing production incidents and system alerts. By integrating into a team's existing observability and ticketing infrastructure, the tool aims to reduce the manual workload associated with on-call duties by investigating, triaging, and, in some cases, remediating issues without immediate human intervention.
According to the company, Devin Oncall is currently in use at various organizations to process thousands of production alerts daily. The system is designed to handle common SRE tasks, such as tracing errors through infrastructure, logs, and codebases to identify root causes. When it encounters issues it cannot resolve independently, it escalates the problem to a human engineer with a summary of its findings and suggested next steps.
Technical Architecture and Capabilities
The engineering behind Devin Oncall is structured into three primary layers: core intelligence, company context, and agent behavior. The core intelligence layer utilizes various frontier models to debug issues within controlled environments. Cognition states that they evaluate these models against benchmarks such as SREGym to determine the most effective model for specific SRE tasks. As models improve, the system's architecture allows for the selection of the most capable model available for a given task.
To prevent "premature conclusions"—a common issue where agents stop at the first plausible explanation—the platform employs an asynchronous "review" agent. This secondary process monitors the primary agent's reasoning, tool calls, and telemetry results. It is designed to flag potential errors, such as misinterpreting correlation as causation or disregarding contradictory data points.
Integration and Workflow
Devin Oncall is designed to be compatible with a wide range of observability and developer tools, including Datadog, Dynatrace, Sentry, New Relic, Grafana, Jira, Linear, and ServiceNow. By connecting to these platforms, the agent can actively monitor channels and incoming alerts. Once triggered, the agent performs a systematic investigation:
- Alert Triage: It correlates incoming signals with logs and traces to identify the underlying cause of an alert.
- Incident Support: It can join incident communication channels, maintain an updated timeline of findings, and assess the potential impact of a system failure.
- Bug Remediation: It is capable of reproducing reported bugs and, when a fix is identified, creating a tested pull request for engineers to review.
A key component of its functionality is its ability to learn from previous incidents. The system maintains a memory of past alerts and their resolutions, allowing it to apply historical context to new issues. For instance, if an alert occurs that matches a previously documented failure, the agent can leverage prior investigation steps to accelerate the current response.
Performance Benchmarking
Cognition highlighted the performance of Devin Oncall on the SREGym benchmark, which tests agents on representative faults derived from real-world production incidents. The company claims that the system has achieved high success rates in identifying root causes and mitigating issues in live systems. In their internal evaluations as of October 2026, the agent demonstrated the ability to handle both identification and mitigation in a single automated run, a task requiring deep integration into the target system's telemetry and deployment workflows.
By automating the initial phases of incident response, the company suggests that teams can minimize "page fatigue" and allow engineers to redirect their focus toward product development, rather than routine debugging and initial triage.
Enjoyed this?
Get more posts like this delivered to your inbox.
