Resolve AI Launches SRE-1 Model and Primary On-Call Service
Production alerts went to an AI agent on Wednesday when Resolve AI launched SRE-1, a 27 billion-parameter model built to find root causes, and Primary On-call, a service that puts its artificial intelligence agents on call ahead of human engineers.
Up to 80% of paged alerts in production are handled by Primary On-call without escalation to a person, according to Resolve AI.
- Resolve AI introduced SRE-1 on November 18, 2026 — a 27B model built on open weights trained with reinforcement learning to investigate production incidents.
- The model is available only through Resolve AI’s platform; Primary On-call is now generally available, with Salesforce Inc. among early users.
- Customers report tenfold improvement in mean time to resolution and root-cause accuracy of up to 95%, figures that come from the company itself.
What 6.9% to 38.8% actually measures
The jump that matters most in the announcement is narrow and honest about its limits. Measured on the same incidents before and after training, which used reinforcement learning alone, a single investigation’s pass rate rose from 6.9% to 38.8%. In other words, the model started barely able to close one investigation and, after RL, could do so a little under four times in ten.
Resolve AI’s own announcement puts a different frame on the same work: on 49 difficult live production cases, each run five times, an ensemble of five SRE-1 investigations matched the strongest frontier model the company tested, at lower cost. That is an ensemble against a single frontier model — a comparison buyers should read carefully before assuming a 27B box beats a frontier system head to head.
System-level results look stronger. On 178 production alert investigations, Resolve’s complete system solved about twice as many incidents as general-purpose coding agents run on the same alerts, per the company’s announcement. On the hardest third of those alerts, its root-cause score was 15 to 25% higher.

| Figure | Value | Comparison the sources give |
|---|---|---|
| Single-investigation pass rate, pre-RL | 6.9% | Same incidents, same tools, before training |
| Single-investigation pass rate, post-RL | 38.8% | Same incidents, same tools, after training |
| Model size | 27B parameters | Built on open weights |
| Training run | 7.4 days, up to 512 GPUs | 16 AI researchers |
| Alerts handled without escalation | Up to 80% | Company-reported, production paging |
| Salesforce dark-launch closure rate | Over 70% | CRM Analytics team paging alerts, first weeks |
Why a domain-specific model at all
Resolve AI’s argument is that AI transformed how software gets written but barely touched how it gets run. The reasoning, laid out in the company’s announcement: code models learned from published lines, while production has no comparable public training corpus. Each organization’s setup is unique, outages are never shared publicly, underlying causes go undocumented, and the data an investigator relies on vanishes within a few weeks.
That last constraint is the technically interesting one. A general-purpose model arrives at an incident without knowledge of the environment it is asked to investigate, and a bigger model does not fix that, because the answer exists only in the running system.
Reaching it means making the right queries, in the right order, and judging what comes back before telemetry ages out. Throughout SRE-1’s training, every tool was frozen at the instant the alert triggered, preventing the model from drawing on information from later points in time.
The build itself is a lean operation: a team of 16 AI researchers, a final training run of 7.4 days on up to 512 graphics processing units. Resolve AI built the model using Snowflake Inc.’s Cortex Training, a serverless service for customizing open-weight models. Snowflake Senior Vice President of Engineering Vivek Raghunathan said Resolve AI was the first design partner for the service.
“Resolve AI’s SRE-1 model shows what a domain-specific model can deliver for production. The fact that Resolve AI has managed to produce this model is a testament to the team’s deep modeling and RL talent. We are proud to have supported Resolve AI as the first design partner for Snowflake Cortex Training, which they used to build SRE-1.”
Vivek Raghunathan, senior vice president of engineering at Snowflake, said that in the company’s announcement.
Primary on-call and the critical-path question
Primary On-call puts that research to work on the critical path. Each page is investigated across telemetry, infrastructure, code and operational knowledge, according to the company. Based on its findings, the agent has three options: close the alert, apply mitigations where team policies allow, or hand the issue up to a person.
Once it has acted, the agent checks the fix against the production baseline and monitors until recovery holds. It also links related alerts and investigations into a shared view of each underlying problem.
For organizations hesitant to give the agent the pager immediately, Resolve AI provides a shadow mode that logs the decision the agent would have reached on every page. The mandate-and-escalate design — close what it can, mitigate only inside policy, escalate with evidence — mirrors how mature on-call policies treat junior engineers.
The economics are the wedge. Chief Executive Spiros Xanthos frames the problem as arithmetic no team can win: “No engineering team can keep up with how fast code and patches land in production. The answer is agents that are accurate and fast enough to sit on the critical path, and that takes a model built for production.”
Salesforce Inc. is among the early users. Bhupesh Patel, a software engineering principal architect at Salesforce, said that during the opening weeks of dark launch testing, the agent resolved over 70% of the CRM Analytics team’s paging alerts “without waking an engineer.”
Who gains, and at what risk
Backed by venture capital, Resolve AI has collected over $190 million so far, among it a $40 million extension of its Series A round in April, led by DST Global and Salesforce Ventures, which put the company’s valuation at $1.5 billion. Its customer roster features Coinbase Global Inc., DoorDash Inc., Figma Inc., MSCI Inc., Nasdaq Inc., SAP SE and Zscaler Inc. SRE-1 itself ships only inside Resolve AI’s platform.
For enterprise buyers, the open questions are operational rather than academic. The 95% root-cause accuracy, tenfold MTTR improvement and 80% no-escalation rate are all company-reported. Teams putting an agent on the critical path still need someone to own blast radius: which actions the agent may execute autonomously, what audit trail persists, and how fast a human can revert a proposed mitigation.
The deployment shape teams will copy
Whatever the accuracy claims turn out to be worth, the rollout sequence is already legible: shadow mode first, then policy-bounded mitigation, then full handoff of the pager. An investigation of this kind is a tool-use loop — query telemetry, check infrastructure state, read the code path, verify against baseline — which is the same shape reliability tooling has always taken, only now driven by a model trained to decide the next query:
Resolve AI was founded by Spiros Xanthos and Mayank Agarwal, observability veterans who co-created OpenTelemetry, the open-source standard for telemetry data, and who each sold earlier companies to Splunk and VMware before leading Splunk’s observability business.
The company emerged from stealth and, 16 months later, had raised more than $150 million in total, including a $125 million Series A at a $1 billion valuation led by Lightspeed Venture Partners, with Greylock Partners, which led the seed round, among the existing investors participating. Its customer roster by then included Coinbase, DoorDash, MongoDB, MSCI, Salesforce, and Zscaler.
# Illustrative: the tool loop an incident agent executes,
# pinned to the moment the alert fired
$ kubectl get pods -n payments --sort-by=.metadata.creationTimestamp
$ kubectl logs payments-api-7f9c4 -n payments --since=15m
$ curl -s http://metrics.internal/payments/latency_p99?window=30m
# compare observed latency against production baseline,
# then close, mitigate (if policy allows) or escalate with findings
Primary On-call is generally available now. But SRE-1 stays locked to Resolve AI’s own platform.