ResolveAI launches incident-response AI that can take the first call before engineers
SRE-1 arrives alongside general availability of Primary On-call. ResolveAI’s performance claims span model tests, whole-system evaluations and customer deployments—three different measures.
ResolveAI is turning incident investigation into a bounded operational service: Primary On-call is generally available, while its SRE-1 model is available only through the platform despite being built on open weights. The service can close alerts, take policy-approved mitigation actions, or escalate with its evidence, and teams can start in shadow mode before assigning selected alert rules. ResolveAI reports positive results on its own evaluations, but the benchmarks measure different things and do not show that one SRE-1 run matches a frontier model or that the claimed 80% alert-handling rate applies broadly.
01
SRE-1 has 27 billion parameters; 16 researchers trained it using Snowflake Cortex Training, with a final run lasting 7.4 days on up to 512 GPUs.
02
Across 49 difficult live cases run five times, ResolveAI says five SRE-1 investigations matched its strongest tested frontier model at lower cost; that is not a single-run comparison.
03
In a separate evaluation of 178 alert investigations, ResolveAI says its complete system solved about twice as many incidents as the general-purpose coding agents tested.
ResolveAI has released an AI model for finding the causes of failures in running software, alongside a service that can receive incident alerts before engineers do. The 27-billion-parameter SRE-1 model supports the company’s push into first-response operations: investigating problems, taking permitted action and calling a person when human judgment is needed.
In its November 18 announcement, ResolveAI made Primary On-call generally available and said SRE-1 would be accessible only through its platform. Although the model was built on open weights—an existing model’s underlying numerical settings—the announced access route is ResolveAI’s service, not a standalone model download. The release pairs the specialized model with a product that can act on its investigations.
ResolveAI’s argument for specialization starts with the evidence an incident requires. The company says private production incidents do not offer a public training resource comparable to published code. Instead, an investigator must query the running environment in the right order and assess the results before useful evidence expires. Its approach trains for that investigative work rather than relying on knowledge learned from the internet.
Training an investigator, not just supplying more context
According to ResolveAI, 16 AI researchers trained SRE-1, with a final run lasting 7.4 days on up to 512 graphics processors. The team used Snowflake Cortex Training. Snowflake engineering executive Vivek Raghunathan identified ResolveAI as the service’s first design partner, describing the model as an example of what specialized training can deliver for production operations.
The training used reinforcement learning, which improves behavior through feedback on attempted tasks. ResolveAI says a single investigation’s pass rate increased on the same incidents with the same tools. To prevent the model from finding answers that emerged after an incident, every training tool was restricted to information available when the alert fired.
For its model comparison, ResolveAI used 49 difficult live production cases, each run five times. It says a group of five SRE-1 investigations matched the strongest frontier model it tested at lower cost. That is a result for combined investigations, not a claim that one SRE-1 attempt matched the competing model.
Editorial illustration for ResolveAI launches incident-response AI that can take the first call before engineers.
A separate evaluation tested the complete ResolveAI system on 178 alert investigations, graded against verified root causes. The company says it solved about twice as many incidents as the general-purpose coding agents tested on those alerts. These are ResolveAI’s own evaluations; the model comparison and system comparison measure different things.
From an investigation to an operational response
Primary On-call investigates an incoming alert across monitoring data, infrastructure, source code and operational knowledge. Its next step depends on the findings and the team’s rules. ResolveAI describes three possible responses, rather than giving the agent unrestricted authority to change whatever it encounters:
Close an alert it can handle without bringing in an engineer.
Mitigate the problem where the team’s policies permit action, then check the result against the production baseline.
Escalate when human judgment is needed, passing along an investigation rather than just the original page.
After acting, the service keeps monitoring until recovery holds. Existing paging-system timeouts remain an independent backstop, and the system records its investigations and decisions for engineers to review. It also connects related alerts to underlying problems, carrying the evidence and investigation history into fixes made within the scope a team has approved.
The pager changes hands one alert rule at a time
ResolveAI says Primary On-call handles up to 80% of paged production alerts without human escalation. Those operational results are separate from the investigation benchmarks. In a testimonial included with the release, Salesforce principal architect Bhupesh Patel said the CRM Analytics team saw more than 70% of paging alerts closed without waking an engineer during its first few weeks of dark-launch testing.
Teams need not hand over the full rotation immediately. ResolveAI describes a shadow stage in which the agent records what it would have done while engineers keep their existing on-call duties. Engineers can compare those decisions with familiar incidents. Primary On-call then takes responsibility only for the alert rules where that record has earned the team’s trust.
Sources
resolve.aiResolveAI Introduces SRE-1, the Most Accurate Model for Production at Every Budget
Reader comments
Newest comments first. Replies stay oldest first.