Resolve AI’s SRE-1 Model Finds Alert Root Causes; RL Boosts Pass Rate to 38.8%

On November 18, 2026, incident remediation startup Resolve AI introduced a 27 billion-parameter model named SRE-1 designed to find the root cause of production alerts, SiliconANGLE reported.

Measured on the same incidents before and after reinforcement learning alone, the pass rate for a single investigation rose from 6.9% to 38.8%.

The company also launched Primary On-call, a new service placing its AI agents on call ahead of human engineers. Resolve AI explains that production environments lack a public training corpus mirroring published code sources, as every company keeps its incidents confidential.

Production lacks a public training corpus to match published code repositories because incidents at each company stay private. SRE-1 was built on open weights by a team of 16 AI researchers over a final run of 7.4 days using up to 512 graphics processing units.

“No engineering team can keep up with how fast code and patches land in production,” said Spiros Xanthos, co-founder and CEO of ResolveAI.

Diagram showing SRE-1 connected to various software and cloud tools
Photo: resolve.ai

Training SRE-1 with Snowflake

Resolve AI utilized Snowflake Inc.’s Cortex Training service to build the new model. Vivek Raghunathan, senior vice president of engineering at Snowflake, said that the startup served as the first design partner for the serverless platform.

In testing across 49 difficult live production cases run five times each, an ensemble of five SRE-1 investigations matched the strongest frontier model tested at a lower cost. Across 178 production alert investigations, the full system resolved roughly twice the volume of incidents managed by standard coding agents.

Deploying Primary on-call

Primary On-call investigates each alert across telemetry, infrastructure, code, and operational knowledge. Depending on its findings, the agent can close the alert, mitigate the problem where policies permit, or escalate to a human engineer.

Agents on-call for you — ResolveAI

Up to 80% of paged alerts in production are handled by the system without escalation. Customers have reported a tenfold improvement in mean time to resolution, alongside root-cause accuracy reaching up to 95%.

Salesforce Inc. tested the agent during a dark launch period. The CRM Analytics team saw more than 70% of its paging alerts successfully closed by the agent without waking up any engineer, according to Salesforce software engineering principal architect Bhupesh Patel.

Common production investigations

What makes SRE-1 different from general-purpose models? General-purpose models arrive at an incident without environment knowledge, whereas SRE-1 is a domain-specific model trained using reinforcement learning alone.

How many parameters does the SRE-1 model have? SRE-1 is a 27 billion-parameter model built on open weights and trained using up to 512 graphics processing units.

Where is SRE-1 available for use? Access to SRE-1 is limited strictly to Resolve AI’s own platform, according to company announcements.

Snowflake Inc. was founded on July 23, 2012, in San Mateo, California, by Benoît Dageville, Thierry Cruanes, and Marcin Żukowski. The American cloud-based artificial intelligence data platform company came out of stealth mode in October 2014, when it was used by 80 organizations, and launched its first product, its cloud data warehouse, in June 2015.

Headquartered in Menlo Park, Snowflake operates a platform that supports data analysis and simultaneous access to data sets with minimal latency, running on Amazon Web Services, Microsoft Azure, and Google Cloud Platform.

Xanthos informed Resolve AI that SRE-1 delivers superior accuracy compared to every other model evaluated at any budget level when handling complex production incidents, including top frontier models.