Breaking
YC W26 Sonarly ships an autonomous AI on-call engineer 78% root-cause accuracy vs 53% for a raw agent 180 → 5 one bad deploy, triaged to the alerts that matter 3 min setup across Sentry, Datadog, Grafana & Slack $0 → $150/mo free to start, startup plan to scale Founders Choudhury & Klobb, ex-Meoria (100k+ users) YC W26 Sonarly ships an autonomous AI on-call engineer 78% root-cause accuracy vs 53% for a raw agent 180 → 5 one bad deploy, triaged to the alerts that matter 3 min setup across Sentry, Datadog, Grafana & Slack $0 → $150/mo free to start, startup plan to scale Founders Choudhury & Klobb, ex-Meoria (100k+ users)
Company · AI & Developer Tools

Sonarly Wants to Answer the 3am Page So You Don't Have To

A two-person Y Combinator W26 team is building an AI engineer that lives inside your monitoring stack - reading logs, tracing failures, and opening the fix before anyone reaches for their phone.

The alert always seems to fire at the worst hour. A deploy goes out, a queue backs up, and somewhere a phone starts buzzing on a nightstand. The engineer on call squints at Sentry, scrolls Datadog, greps a log, and eventually - bleary, annoyed - finds the one line that mattered buried under a hundred that didn't. Sonarly was built by two people who lived that routine enough times to decide a machine should do it instead.

Sonarly is a San Francisco company in Y Combinator's Winter 2026 batch, and its pitch is unusually literal: an AI engineer for production. It connects to the monitoring tools a team already runs, watches for alerts, and when one fires it does the work a human on-call would do - triage the noise, investigate the cause across logs, traces, metrics and code, and open a pull request with the fix. The company describes the goal as making software self-healing. The less poetic version is that it wants to shrink the time between "something broke" and "something is fixed."

78%
Root-cause accuracy reported
3 min
Self-serve setup time
2
People on the team
The problem

01The bug you didn't write, at the hour you didn't want

Founders Dimittri Choudhury and Alexandre Klobb didn't arrive at this from theory. Their previous company, Meoria - an app they've described as "the Duolingo for finding schools" - reached more than 100,000 users in about six months and, in their telling, made them drop out to chase it. Running a consumer product at that scale meant living with the exhaust of production software: on some days, up to fifty Sentry alerts, most of them noise, a few of them real, and no fast way to tell which was which.

That is the quiet tax on most engineering teams. Alert fatigue is usually framed as a productivity problem, but it behaves more like a retention problem. When the signal is buried, real bugs live longer, and users feel the broken product before anyone gets around to the fix. Sonarly's founders decided the interesting work wasn't a prettier dashboard. It was removing the human from the boring middle of the loop entirely.

"You can't simply pipe every alert directly into an agent. Effective fixes require deep runtime context and intelligent triage." Sonarly, Launch HN
How it works

02Triage first, then let the agent loose

The temptation with any coding agent is to wire it straight to your alerts and hope. Sonarly's whole argument is that this doesn't work - and they have a number for it. A raw agent pointed at monitoring tools through generic connections lands the root cause about 53% of the time. Add Sonarly's triage layer and a running map of the system, and that figure climbs to 78%. The difference isn't a smarter model. It's the preparation that happens before the model reads a single line.

1Ingest - alert arrives via webhook from Sentry, Datadog or Grafana
2Dedupe - cluster duplicates, drop false positives
3Context - pull logs, traces, metrics and user feedback
4Investigate - a coding agent finds the root cause
5Score - rate severity and confidence to avoid guesses
6Fix & learn - open a PR, then update the system map
The Sonarly loop: the noisy top of the funnel is handled before the agent - powered by Claude - is ever asked to reason.

Underneath sits what the founders call a living map: a dynamically generated markdown model of a customer's production system, tracking services, dependencies, and the history of what has broken before. It is, effectively, the institutional memory an on-call engineer builds over months, handed to the agent on day one. When an alert fires, Sonarly doesn't start from scratch. It starts from a picture of how everything fits together.

Sonarly (triage + living map)78%
78%
Raw agent + generic monitoring hookup53%
53%
Root-cause accuracy, as reported by Sonarly. The gap is context, not model size.
The math of noise

03From a wall of alerts to a short list

The clearest way to understand what Sonarly does is to watch a bad deploy move through it. One customer, the company says, saw roughly 180 alerts from a single deploy collapse to about 50 unique issues after deduplication, then narrow to around 5 that were actually worth a human's attention. That is the part on-call engineers rarely get credit for and never enjoy: separating the five that matter from the hundred and seventy-five that don't.

~180 raw alerts from one bad deploy
~50 unique issues after dedup
~5 actually actionable
A reported customer example. Most of on-call is subtraction; Sonarly does the subtracting.
Where it fits

04The new race to automate operations

Sonarly launches into a crowded and fast-moving corner of the market. There are AI-SRE startups like Cleric and Resolve AI aiming at the same on-call pain, and there is the do-it-yourself route of pointing a general coding agent at your monitoring tools. Sonarly's bet is that the winners won't be whoever has the flashiest agent, but whoever does the unglamorous work of triage and context best. In a field where everyone can call the same underlying models, the differentiator is what you feed them.

The business model is familiar developer SaaS. There's a free tier to get teams onboarded, a startup plan around $150 a month, and custom enterprise pricing, with a 14-day trial. Integrations - Sentry, Datadog, Grafana, Slack, Discord, Linear, GitHub, PagerDuty - connect through OAuth, and the setup is pitched at three minutes. The trial-and-expand shape is deliberate: land on the free plan, prove the fixes are real, grow with the team.

Most of on-call isn't fixing. It's deciding what deserves fixing. Sonarly is a bet that a machine can make that call - and then keep going.
The people

05Second-time founders, same instinct

Choudhury, who serves as CEO, moved from a small village in France to Paris to study computer science before the pair ended up in San Francisco for YC. Klobb brings the full-stack and ML engineering side. Their throughline is consistent: build things people actually use, ship quickly, and treat user feedback as the roadmap. It's the same "make something people want" instinct that carried Meoria to six figures of users, now pointed at a problem they know from the inside. The company operates under the parent name Mirabelle - a small yellow plum from the Lorraine region, a quiet nod home.

Whether an AI can genuinely own an on-call rotation is still an open question, and Sonarly is early - two people, a young batch, a handful of design partners. But the framing is sharp. Nobody wants "an agent." Everybody wants the 3am page to stop and the fix to be waiting in review by morning. Sonarly is selling the second thing, and hoping the first thing becomes true along the way.


Watch: no official product-demo video was publicly available at publication. Check sonarly.com and the YC launch page for the latest walkthroughs.

#ai#developer-tools#on-call#observability #sre#incident-response#root-cause-analysis #sentry#datadog#yc-w26#mttr#saas