ResolveAI introduced the SRE-1 model on November 18, 2026, a 27-billion-parameter system built to trace production alerts back to their source. Alongside it, the San Francisco company launched Primary On-call, an agent that answers pages first and wakes an engineer only when it has to.
SRE-1 is available only through ResolveAI. It is not sold as a standalone model.
How the SRE-1 model was trained
SRE-1 is built on open weights. Sixteen AI researchers ran the last training pass over 7.4 days, using as many as 512 GPUs. Training used reinforcement learning alone.
The test set was 49 hard, live production cases, each attempted five times. With identical tools and incidents, one investigation’s pass rate climbed from 6.9% to 38.8%. Combining five SRE-1 investigations matches the best frontier model ResolveAI tried, for less money.
One design choice stands out. In training, each tool was frozen at the instant the alert fired, which kept the model from seeing how things turned out. Without that, a model could read the answer from the future and look better than it is.
ResolveAI built the model with Snowflake Cortex Training, and Snowflake says ResolveAI was its first design partner for that service.
Why a general model struggles with incidents
ResolveAI’s argument is that code has a public training corpus and production does not. Every company runs a different system, incidents stay private, causes rarely get documented, and the evidence disappears in a few weeks.
A bigger general model doesn’t fix that, the company says, because the answer exists only in the running system. Reaching it means asking the system the right questions in sequence and weighing the replies while the evidence still exists. In ResolveAI’s view, that skill has to be trained into a model.
Benchmarks for the full system
The company also measured the whole product, not only the model. Across 178 production alert investigations, scored against confirmed root causes, ResolveAI resolved roughly double the incidents that general coding agents did. The gap in root-cause score was 15 to 25% on the toughest third.
These are the vendor’s own tests. The company doesn’t name the competing coding agents in its announcement, and outside teams haven’t reproduced the results.
What Primary On-call does
When a page arrives, Primary On-call digs through telemetry, infrastructure, code and the team’s operational knowledge. Then it either closes the alert, applies a mitigation if team policy permits, or hands the case to a person along with its findings. Once it acts, it compares the result with normal production behavior and stays on watch until the recovery sticks.
ResolveAI reports these figures from production use:
- Up to 80% of paged alerts handled without escalating to a human.
- Customers report mean time to resolution ten times better.
- Root-cause accuracy of up to 95%.
The timeouts already in paging systems remain as a separate safety net, and every investigation and decision is recorded so engineers can review what the agent found and why it acted.
Salesforce’s CRM Analytics team said that during the first weeks of a dark launch, ResolveAI closed more than 70% of its paging alerts without waking an engineer. Bhupesh Patel, a software engineering principal architect there, said the team is ready to put Primary On-call in its paging path as first responder. Zscaler’s chief reliability officer, Misha Kuperman, also backed the approach.
Fixing the cause, not only the page
If a page is answered but the cause survives, it will fire again, and so will every other alert that cause generates. ResolveAI says its agents also go after those root problems, which removes a source of future incidents. CEO Spiros Xanthos says code and patches now land faster than any engineering team can follow.
What to watch next
The claims that matter most are the 80% handling rate and the root-cause accuracy, since both come from the company and its customers. Teams considering Primary On-call will want to see how it behaves on their own systems, and how often its closed alerts come back.
The test of the SRE-1 model is whether on-call rotations shrink without more missed incidents. ResolveAI says uptime should no longer hinge on staffing a large on-call rotation. That is a bold promise, and the first real checks will come from customers who now trust it with the pager.
Source: ResolveAI