Resolve AI’s new 27-billion-parameter SRE-1 model lifted a single investigation’s pass rate from 6.9% to 38.8% through reinforcement learning alone, and the company is now putting the agents built on it directly on call.
The startup introduced SRE-1 and its Primary On-call service on November 18, betting that AI should run software, not just write it.
That gap between writing and running software is the core of Resolve AI’s argument. Models learned to code from billions of published lines. Production offers no such corpus: incidents stay private, root causes are never written down, and the telemetry evidence an investigation depends on expires within weeks.
A bigger general-purpose model doesn’t fix that, because the answer was never on the internet to begin with. It exists only in the running system, reachable by querying it in the right order before the traces age out.
So Resolve AI trained a domain-specific model instead. SRE-1 is a 27B model built on open weights, trained by 16 AI researchers over a final run of 7.4 days on up to 512 GPUs. The company built it on Snowflake’s Cortex Training, a serverless service for customizing open-weight models, and Snowflake says Resolve AI was that service’s first design partner.

The numbers tell a two-sided story. On the same incidents, before and after training, reinforcement learning took a single investigation’s pass rate from 6.9% to 38.8%. That is a dramatic improvement — and still under 40% on its own. Resolve AI’s answer to that ceiling is the ensemble: across 49 difficult live production cases, each run five times, five SRE-1 investigations working together matched the strongest frontier model the company tested, at lower cost.
On 178 production alert investigations graded against verified root causes, the complete system solved about twice as many incidents as general-purpose coding agents, with the widest margin on the hardest third of alerts, where the company’s root-cause score ran 15 to 25% higher.
Why production defeats general-purpose models
The training discipline matters as much as the architecture. Each tool was frozen at the timestamp of the alert’s firing during training, blocking the model from peeking at information that came later — a safeguard Resolve AI’s announcement itself calls out, and one that distinguishes a legitimate benchmark from one that leaks ground truth.
General-purpose coding agents arrive at an incident knowing nothing about the environment they are investigating. SRE-1 was trained to make the right queries, in the right sequence, and judge the results before the evidence expires.
“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. On hard production incidents, SRE-1 is the most accurate model at every budget we measured, including frontier models.”
That is Spiros Xanthos, co-founder and CEO of ResolveAI, in the company’s announcement. The claim worth watching is “at every budget we measured” — Resolve AI frames SRE-1 as dominating the Pareto frontier of pass rate against cost, meaning no tested model beat it on accuracy without paying more.
Primary on-call takes the pager
SRE-1 stays locked to Resolve AI’s own platform. What ships broadly is Primary On-call, now generally available. Each page is investigated across telemetry, infrastructure, code and operational knowledge. Depending on what it finds and what a team’s policies permit, the agent can close the alert, mitigate the problem, or escalate to a human with an investigation attached.
After acting, it checks the fix against the production baseline and monitors until recovery holds. It also links related alerts into a shared view of each underlying problem — going after the issues that keep generating pages, not just the pages themselves.
For teams unwilling to hand over the pager on day one, there is a shadow mode that records what the agent would have done on each page without acting. Resolve AI says up to 80% of paged alerts are handled without escalation, and customers report a tenfold improvement in mean time to resolution with root-cause accuracy of up to 95%. Those are the company’s own figures, from its own benchmarks.
Salesforce is among the early users. During the opening weeks of dark launch testing, the agent resolved over 70% of the paging alerts belonging to the CRM Analytics team, according to Bhupesh Patel, a software engineering principal architect at Salesforce, “without waking an engineer.”
The open-weights bet on a closed platform
There is a tension here worth naming. SRE-1 is built on open weights — the kind of foundation that usually feeds an ecosystem of third-party deployments — yet access is limited to Resolve AI’s platform only.
The company has raised more than $190 million to date, including a $40 million Series A extension in April led by DST Global and Salesforce Ventures at a $1.5 billion valuation, and its customer list runs from Coinbase and DoorDash to Nasdaq and SAP. Keeping the model proprietary is how a startup at that valuation defends the moat; open weights are simply the most efficient training substrate, not an open-source pledge.
The contrast with the broader agent market is sharp. General-purpose coding agents are being repurposed for operations on the assumption that scale substitutes for domain knowledge. Resolve AI’s benchmarks, as the company presents them, say the opposite: the widest margin over coding agents came on the hardest alerts.
If that holds up under independent scrutiny, domain-specific models for operations become a category rather than a curiosity — and every observability vendor without one has a gap to explain.
What is not yet known is how these numbers survive outside Resolve AI’s own grading. The benchmarks are the company’s, the incidents are private, and the verified root causes are its own. Shadow mode exists precisely to let skeptical buyers run that test on their own pager before committing.
Snowflake, the company whose Cortex Training service Resolve AI used to build SRE-1, was founded on July 23, 2012, in San Mateo, California, by Benoît Dageville and Thierry Cruanes, both former data architects at Oracle, together with Marcin Żukowski.
It spent its first two years in stealth mode, emerging in October 2014 with 80 organizations already using it. The Cortex generative AI services that include Cortex Training were launched in 2023, and in February 2026 Snowflake reported full-year product revenue of $4.72 billion, with over 9,100 accounts using its AI features.
The details are on Resolve AI’s announcement of SRE-1 and Primary On-call. SRE-1 remains available only through the company’s platform; Primary On-call is shipping now.