ResolveAI SRE-1 Model Puts AI Agents On Call
ResolveAI on Nov. 18 released SRE-1, a 27B open-weight-based model trained with reinforcement learning to find root causes of production alerts, and made its Primary On-call agent generally available. The company says five combined SRE-1 runs matched the best frontier model it tested at lower cost; the figures are self-reported.
ResolveAI on Nov. 18, 2026, released SRE-1, a 27-billion-parameter model trained to find the root cause of production alerts, and made generally available Primary On-call, a service that lets its AI agents take the pager before human engineers do. The San Francisco startup says SRE-1 is available only inside its own platform.

Key takeaways
- SRE-1 is a 27B model built on open weights and trained with reinforcement learning alone, which raised a single investigation's pass rate from 6.9% to 38.8% on the same incidents.
- On 49 hard live production cases, an ensemble of five SRE-1 investigations matched the strongest frontier model ResolveAI tested at lower cost.
- Primary On-call is generally available; ResolveAI says it handles up to 80% of paged alerts without escalating to a person.
What is ResolveAI SRE-1?
ResolveAI argues that AI has changed how software is written but barely touched how it is run. Code models learned from billions of public lines, while production incidents stay private, root causes are rarely written down and telemetry expires within weeks, the company said in its announcement.
A team of 16 AI researchers trained SRE-1 over a final run of 7.4 days on up to 512 GPUs, using Snowflake's Cortex Training service. Every tool used in training was pinned to the moment an alert fired, so the model could not read the answer from later commits or tickets. Snowflake engineering SVP Vivek Raghunathan said ResolveAI was the first design partner for Cortex Training.
SRE-1 benchmark results vs. frontier models
ResolveAI counts a run as a pass when LLM judges score the model's causal chain at 0.8 or higher against a reference chain, a threshold it calibrated with on-call engineers. On 49 difficult live cases, each run five times, five combined SRE-1 investigations matched the best frontier model tested. The company's technical report page names that model as Fable 5.2 and puts the ensemble's cost at 0.71 times as much. A single SRE-1 run does not make that claim.
The full ResolveAI system was tested separately on 178 production alert investigations graded against verified root causes. It solved about twice as many as general-purpose coding agents run on the same alerts, and scored 15% to 25% higher on root cause for the hardest third. All results are self-reported and have not been independently reproduced.
How Primary On-call works
For each page, the agent searches telemetry, infrastructure, code and operational notes, then closes the alert, mitigates where a team's policy allows or escalates with its investigation attached. Existing paging timeouts stay in place as a backstop, and teams can first run a shadow mode that records the call the agent would have made.
"No engineering team can keep up with how fast code and patches land in production," said Spiros Xanthos, ResolveAI's co-founder and CEO. Bhupesh Patel, a principal architect at Salesforce, said that in the first weeks of dark launch testing, the agent closed more than 70% of his CRM Analytics team's paging alerts "without waking an engineer."
Context: AI SRE agents move onto the critical path
ResolveAI was founded by Xanthos and Mayank Agarwal, co-creators of OpenTelemetry, and lists Coinbase, DoorDash, Figma, Nasdaq, SAP and Zscaler among its customers. According to SiliconANGLE, it has raised more than $190 million, including an April extension that valued it at $1.5 billion.
The launch comes days after Microsoft expanded its own Azure SRE Agent at Ignite, as cloud vendors and startups compete to let agents act in production. Giving agents write access carries risk; researchers have documented agents that turned to SQL injection when ordinary requests failed.
What's next
ResolveAI plans to show the system at AWS re:Invent, which runs Nov. 30 to Dec. 4. Because SRE-1's weights are not released, outside teams cannot yet test the benchmark claims on their own incidents.
FAQ
Can developers download SRE-1? No. It is built on an open-weight base model, but ResolveAI offers it only through its own platform.
Does Primary On-call replace human on-call engineers? Not entirely. It escalates when human judgment is needed, and existing paging timeouts remain as a backstop.
Are the accuracy figures independent? No. The benchmark and customer numbers come from ResolveAI and its customers.
References
- ResolveAI Introduces SRE-1, the Most Accurate Model for Production at Every Budget — ResolveAI (Nov. 18, 2026)
- SRE-1: Frontier Incident Investigation from a 27B Model — ResolveAI Labs (Nov. 18, 2026)
- Resolve AI's SRE-1 model finds root causes of production alerts as AI agents go on call — SiliconANGLE (Nov. 19, 2026)
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.