ResolveAI has trained its own model to find out why production systems break. On 18 November 2026 the San Francisco company, founded by OpenTelemetry co-creators Spiros Xanthos and Mayank Agarwal, introduced SRE-1, a 27B model built on open weights and trained with reinforcement learning for root cause analysis, and made its Primary On-call agent generally available.
Quick Take
- The model: SRE-1, a 27B dense model post-trained with reinforcement learning only, available only through ResolveAI.
- The jump: RL took a single investigation’s pass rate from 6.9% to 38.8% on the same incidents with the same tools, according to the company.
- Cost claim: one SRE-1 investigation matches the strongest frontier model ResolveAI tested at medium effort for about a fifth of the cost.
- The agent: Primary On-call takes the page, closes what it can and escalates to engineers with an investigation attached.
- Training: a final run of 7.4 days on up to 512 GPUs by a team of 16 researchers, using Snowflake Cortex Training.
In plain English
When a website or app breaks at 3 a.m., an engineer gets paged and has to dig through logs and dashboards to find the cause. ResolveAI’s agent now takes that page first. Its new model is trained specifically for this detective work, and the company says it often fixes or closes the alert without waking anyone.
What happened
ResolveAI announced SRE-1 and Primary On-call in a press release, a company blog post and a technical paper. “No engineering team can keep up with how fast code and patches land in production. The only answer is agents accurate, fast, and cheap enough to put on the critical path, and that takes a model built for production,” said Spiros Xanthos, co-founder and CEO.
How it works
- Training environments: live incidents and replayed telemetry from ResolveAI’s own production, synthetic systems built by agent teams, licensed vendor systems and offline synthesis from saved evidence.
- No peeking: every tool in training was pinned to the moment the alert fired, so the model could not read the fix from later commits or tickets.
- Output: a cited causal chain from root cause to the alert, scored by LLM judges against a reference chain.
- Ensembles: five SRE-1 investigations combined match the frontier model at its highest effort for 37% less, the company says.
- System benchmark: on 178 production alert investigations, ResolveAI says its system solved about twice as many incidents as general-purpose coding agents on the same alerts.
Why it matters
For engineering teams, ResolveAI claims Primary On-call handles up to 80% of paged alerts without escalation in production. Salesforce’s Bhupesh Patel said that in dark launch testing it “closed more than 70% of our paging alerts without waking an engineer.”
For the AI industry, SRE-1 is another example of a small, domain-specific model beating general frontier models on price-performance, like Microsoft’s routing model for agent tasks.
For companies, putting agents on the critical path raises the stakes: ResolveAI keeps existing paging timeouts as a backstop and records every decision for review.
What’s still open
- The base model behind SRE-1 and whether its weights will ever be released.
- Independent evaluations; all accuracy figures come from ResolveAI’s own tests.
- How autonomous remediation is limited in regulated environments.
Sources
- ResolveAI, “ResolveAI Launches SRE-1, the Most Accurate Model for Production at Every Budget,” press release, 18 November 2026
- ResolveAI blog, “Agents on the critical path is the only viable way forward to keep up with software change,” 18 November 2026
- ResolveAI Labs, SRE-1 overview and technical paper, 18 November 2026
i10x Perspective
Coding agents made writing software cheaper; running it is the next bottleneck. ResolveAI’s argument that production knowledge cannot be scraped from the internet is a strong case for specialised models. The proof will be fewer 3 a.m. pages at customers, not benchmark charts. More in our Agents section.



