Cognition has launched Devin Oncall, an AI site reliability engineer that investigates alerts and bugs on its own. The company announced it on X on October 15, 2026: "Devin automatically investigates alerts and bugs, helping your team diagnose and resolve production incidents. Devin Oncall already handles thousands of alerts every day for some of the world's largest companies. Let it become your SRE today."
The product is real; the headline number is a claim. This post separates the two, explains what Oncall builds on, and gives a way to test any AI on-call agent before you trust it with a pager.
Quick reference
| Question | Answer |
|---|---|
| What is it? | Devin Oncall, an AI SRE that investigates production alerts and bugs |
| Announced | October 15, 2026, by Cognition on X |
| Headline claim | Thousands of alerts handled every day for some of the world's largest companies |
| Verified? | Not independently. No customers, resolution rates or accuracy figures named in what we found |
| What it builds on | Devin Auto-Triage (May 2026), Slack, Linear, GitHub and webhook triggers |
| Pricing | Not stated in the announcement we could read |
| Key risk | Untrusted alert text, wrong-but-confident diagnoses, skill atrophy |
What Cognition actually said
Cognition's pinned post is short, and we could not find a longer launch article on its blog at the time of writing; the blog index we read ended in late September. So the only first-party words are the post above. We did not find pricing, supported integrations, or a list of customers for Oncall itself.
What we can read is the product it grows out of. Cognition's Auto-Triage announcement from May 18 says Devin can "monitor incoming alerts, investigate them automatically, and come back with an investigation, next steps, and even a PR." It lists triggers such as Slack messages, Linear events, GitHub activity, schedules and webhooks, and says Devin can inspect the codebase, check observability tools, look through related tickets, ask for missing context and spin up sub-Devins to investigate in parallel. If it finds a likely cause it posts a summary; if a human is needed it tags the owner; if the fix is clear it can open a pull request.
The post also describes memory: Devin keeps context about prior investigations and routing preferences, and can connect a known issue's new alerts to the earlier thread, de-duplicating incidents. The only named customer quote is from Modal's inference team, which said Devin monitors its channel without prompting and returns fixes or next steps.
The thousands-of-alerts claim, checked
"Thousands of alerts a day" tells you volume, not value. A reader needs several other numbers:
- How many alerts were actually resolved, not just touched.
- How often the diagnosis was right, and how that was judged.
- Time to first useful finding compared with a human on call.
- How many pull requests were merged without edits.
- How many wrong diagnoses a human acted on.
None is in the material we found. This is the same pattern we flagged when Cognition cut costs with Fusion: the company publishes benchmarks and headline figures, and independent confirmation lags. That does not make the claim false. It means you should ask for the denominator before you rely on it.
What Cognition has been building toward this
Oncall lands after a run of moves that make more sense as a set:
- In July Cognition said it was welcoming TierZero's team, saying they had "gone deep on everything that keeps software running once it ships" and would bring their automations work into Devin.
- In May it launched Auto-Triage, and in September it announced a multi-year AWS collaboration aimed at getting autonomous engineers into production.
- It has released new models like SWE-2 and a memory feature covered in our post on Devin's dreaming memory, which matters because on-call work depends on remembering earlier incidents.
- It also acquired the team behind Poke in July, covered in our Poke acquisition post, a sign that it wants Devin reachable in chat where incidents are handled.
Together this reads as a move from "agent that writes code" to "agent that keeps running code healthy," where the money is in avoided downtime and saved engineer hours.
Where it sits in a crowded AI SRE market
Cognition is not alone. Datadog announced Bits AI SRE in December 2025 as an on-call agent that investigates alerts and surfaces the fastest path to remediation, and observability and delivery vendors such as Harness describe similar AI SRE products. The difference Cognition is betting on is that Devin already writes and tests code in a virtual machine, so the step from "here is the likely cause" to "here is a pull request that fixes it" is short. Observability vendors start from the data and add action; Cognition starts from the code agent and adds the data connections.
For a buyer, that suggests different questions. With an observability vendor, ask how well it reads your telemetry. With a code-agent vendor, ask how well it reads your telemetry and how safe it is to let it write. Neither claim is verified by anything published today.
What an incident looks like with an agent in the loop
Cognition's Auto-Triage description gives the shape of the workflow, and Oncall is positioned to run it continuously. Suppose a latency alert fires at night. The agent receives the alert from a webhook or channel, pulls recent deploys, checks dashboards and logs through connected tools, looks at related tickets, and groups it with earlier alerts if it matches a known issue. It then posts a summary with a suspected cause, tags the owner, and, if the fix is clear, opens a pull request.
Each step is a place to check. Did it look at the right deploy? Did it cite the log lines it relied on? Did it skip a dependency owned by another team? A good on-call agent makes these steps visible, so a human can audit the reasoning in thirty seconds instead of redoing the investigation. A bad one gives a paragraph of confident prose with no trail. When you evaluate Oncall, insist on the trail.
What this means for what you build or pay
If you run production systems, the immediate question is whether the cost of an agent per alert is lower than the cost of engineer attention per alert, including the review time. Cognition has not published pricing for Oncall, so you cannot do that calculation yet; ask for it in writing, along with how usage is metered during an alert storm, when thousands of related alerts could otherwise multiply cost.
If you build agents, the lesson is about inputs. An on-call agent reads text written by systems and customers. Design for the case where some of that text is hostile, keep write permissions narrow, and log every tool call.
Risks to weigh before you give it a pager
Prompt injection through alerts. Alert payloads, logs and tickets are text anyone can influence. Cognition's Auto-Triage post acknowledges this: such inputs "can contain arbitrary text and should not be treated as trusted instructions," and says Devin runs in network-sandboxed environments with extra protections against prompt injection and data exfiltration. Treat that as a design claim to test, not a guarantee. Our coverage of a GitHub agentic-workflow prompt injection bug shows how untrusted text reaches agents.
Confident wrong answers. An investigation that reads well can send a tired human down the wrong path at 3 a.m. Ask for evidence links, not just conclusions.
Skill atrophy. If the agent handles most incidents, humans see fewer of them. Our post on AI SRE and incident-response skill atrophy covers why that matters when the agent is wrong.
Blast radius. Decide what Devin may do: read logs, comment, open a PR, restart a service. Start with read-only plus comments.
If you are connecting agents to production systems, the explainx.ai team's AgentBeam stops AI agents before they take dangerous actions, which is the kind of guardrail an on-call agent needs around write access.
How to test an AI SRE before you trust it
- Replay past incidents. Feed it a month of historical alerts and compare its diagnosis with the postmortems.
- Track precision. For each alert, record whether the root cause it named was right, partly right or wrong.
- Measure time saved honestly, including the time humans spend verifying.
- Test with hostile input. Put an instruction in a log line and see whether the agent obeys it.
- Limit permissions. Give read-only access first; require human approval for any change.
- Keep humans in rotation. Run game days without the agent so the team keeps the skill.
What to watch
- A full Cognition launch post with pricing, integrations and metrics.
- Named customers and measured outcomes.
- Independent evaluations against competing AI SRE products.
- Whether Oncall is bundled into existing Devin plans or priced per alert.
Related reading
- Cognition and AWS strategic collaboration
- Devin Fusion and the 39 percent efficiency claim
- SWE-2 from Cognition
- Devin's dreaming memory
- AI SRE and skill atrophy
- Cognition acquires Poke
Details are accurate as of October 15, 2026, based on Cognition's X post and its Auto-Triage blog post. Claims about scale are Cognition's own.