How Devin root-causes 70% of incidents at AngelList before an engineer steps in

Vimeo
3.5x
More incidents root-caused on first pass than with their homegrown agent, from 20% to 70%
50%
Incidents handled by Devin end-to-end in a recent week
~20
Incidents a week flowing through Devin Oncall

About the company

AngelList is one of the largest fund-administration platforms in venture capital, with $171B in assets on platform across 25,000 funds and syndicates.

Industry: FinTechVisit site

Building their own on-call agent first

AngelList's on-call engineer sees around 20 incidents a week through incident.io, from low-severity issues to full outages.

In early 2026 the team built its own investigation harness on Claude Agent SDK, called Sky, to take a first pass at each one. Sky ran for about six months. It was useful, but often noisy in an incident channel, and its suggestions were often wrong or hallucinated. About one in five of its proposed root causes held up. Engineers were spending their on-call time verifying the agent's guesses.

Testing Devin Oncall against their homegrown agent

In August, AngelList turned on Devin Oncall as an experiment and ran the two systems in parallel on the same incidents. Devin is added automatically to every incident channel that incident.io creates. It makes a first investigation, posts relevant links, decides whether to accept the incident or decline it as a false alarm, then forms hypotheses about the root cause and tests them against the codebase and the running system. The team also pointed Devin Oncall at their Sentry alerts.

"Coming up with hypotheses and testing them is something that really shines with Devin Oncall."

— Thibaut LaBarre, Engineering Lead, AngelList

One incident particularly impressed the team. An issue around customer emails not going out had surfaced months earlier, and a senior engineer investigating it at the time had been stumped. When it recurred, the Claude Code harness proposed a root cause that turned out to be hallucinated. Devin found the actual cause and produced both an immediate fix and a longer-term one so the incident couldn't come back. The team ran an internal comparison of Devin and Sky's findings against human-verified investigations, and found Devin was right more often.

Making Devin the primary responder

About two weeks after the comparison, the team turned Sky off and made Devin Oncall the main incident responder. Everyone on the engineering team was on board with the switch. Devin now correctly identifies the root cause on roughly 70% of incidents, up from about 20% with Sky, so the on-call engineer can skip the investigation and go straight to remediation. In a recent on-call week, Devin handled about half the incidents on its own. For the rest, the engineer picks up from Devin's findings and asks it to open a pull request, run a blast-radius analysis, or identify a console script to run. Sky posted constantly; Devin chimes in when it has something to add, which made it easier to leave it working in the channel.

"It feels like we got an upgrade in our sidekick. We now have a very smart agent taking a first pass at the problems."

— Thibaut LaBarre, Engineering Lead, AngelList

Getting ahead of the incident curve

Because the first proposed root cause is now usually correct, there is less back and forth per incident, and the on-call engineer has time left over to fix root causes and do incident-reduction work that previously kept getting pushed aside. The two weeks since the full switch have been among AngelList's lowest weeks for time spent on incidents.

Asked what it would mean to lose Devin Oncall, Thibaut didn't hesitate.

"I'd be extremely sad. I don't like being on call. It would be a massive step back."

— Thibaut LaBarre, Engineering Lead, AngelList