Raindrop Catches Agent Failures Live

Most evals catch the bugs you already imagined. Raindrop is selling the other problem: the agent that fails politely, at scale, until a human notices.

On 17 September 2026 the San Francisco company said it had closed a Series A led by CRV. Total funding is $50 million. It did not publish the size of this round. Lightspeed and Y Combinator came back. So did personal checks from researchers and executives at Anthropic, OpenAI, and Thinking Machines. Figma Ventures and Vercel Ventures are on the list. Named customers include Vercel, Framer, and Clay, plus unnamed Fortune 100 teams in healthcare and logistics.

Image: Raindrop, Series A announcement, 17 September 2026.

What it watches in production

Raindrop reads live traces. It flags what it calls semantic anomalies: hallucinated answers, tools used the wrong way, and behavior that shifts after a model upgrade. The dashboard is supposed to show what changed, when it started, and which users got hit.

CEO Zubin Koticha's line is the category pitch: agents now run for hours, call thousands of tools, and touch money, health data, and customers. When they fail, they do the wrong thing convincingly.

Simulations: test the change before it ships

The new product, Raindrop Simulations, is in early access and due for general availability over the coming month. It runs on every pull request. It replays real production traffic and your existing tests against a proposed change to the agent harness, then runs the same anomaly detector on the result.

The hard part, they say, is harness-agnostic replay. You cannot just replay cached tool responses. Add a new tool and the old trace has no hole for it. Move the harness to Pi and the world around the agent has to be rebuilt. Their metaphor is Swiss cheese: each tool call is a hole that shows the real state of the world. That is what they reconstruct.

They point at lab practice: OpenAI's deployment simulation (regenerate answers to de-identified production chats with a candidate model) and Anthropic's synthetic universes. The claim is that companies outside the labs now get a version of that loop. Simulations is still preview. The claim is not settled.

Where this sits

CRV's Reid Christian frames agent failure as a detection problem, the way a security company would. That is the opposite of "write better evals." It is also crowded: groundcover, Scaled Cognition, Harvey buying Guardrails AI. Raindrop's bet is the pull request, plus production, not a yearly certificate.

Founders Koticha, Ben Hylak, and Alexis Gauba hired people who built fraud models at Robinhood and anomaly detection at Square. If you already ship agents, the buying question is simple: can it show a silent fail from last Tuesday, with the user list, and can Simulations fail a PR for a behavior you never wrote a test for? If both answers are no, you bought another dashboard.

Sources

Raindrop, Announcing our Series A and $50M in total funding, 17 September 2026. https://www.raindrop.ai/blog/series-a/

The Next Web, Raindrop hits $50m in funding to catch AI agents failing in production, 18 September 2026. https://thenextweb.com/news/raindrop-series-a-50m-crv-agent-failures-simulations