Raindrop raises $35M and launches pre-release testing for AI agents
CRV led the round, bringing total funding to $50M, as Raindrop moves from detecting production failures to testing agent changes before release.
By Ryan Merket · Published
Primary source: X
Why it matters
Raindrop is turning production failures into reusable pre-release tests, a position that could let it own the feedback loop between agent monitoring, debugging, and deployment.
Ben Hylak (@benhylak), Zubin Singh Koticha (@snarkyzk), and Alexis Gauba (@alexisgauba) have raised a $35 million Series A for Raindrop, pairing the financing with a product designed to catch AI agent failures before code reaches production.
https://x.com/benhylak/status/2100630323382088033
CRV led the round, according to Axios. The financing brings Raindrop's total capital raised to $50 million, following a $15 million seed round led by Lightspeed Venture Partners in December 2025. Raindrop said in its September 17th announcement that researchers and executives at Anthropic, OpenAI, and Thinking Machines also backed Raindrop.
The simultaneous funding and product launch show where Hylak, Koticha, and Gauba intend to take Raindrop. Raindrop began by monitoring agents after deployment and identifying recurring failures across production traces. Its new Simulations product moves that inspection into the software development cycle, where a broken agent can be stopped before customers encounter it.
From production alarms to pre-merge tests
Raindrop Simulations runs an agent's real code against synthetic versions of the services it uses, including databases, communication tools, payment systems, and external APIs. Raindrop says the system can replay relevant production traces for each pull request, detect behavior changes, and post a verdict before engineers merge the code.
That approach targets a weakness in conventional agent evaluations. Handwritten tests cover failures that engineers already expect. Agents can also break through an unanticipated sequence of tool calls, stale context, a changed prompt, or a model substitution that alters behavior several steps into a task.
Raindrop says Simulations uses production traffic and existing tests, reruns them against a proposed agent change, and applies its anomaly-detection system to the resulting trajectories. The product then reports regressions, cost changes, output drift, and tool errors in the pull request.
The difficult technical work sits in reproducing the world surrounding the agent. Replaying a cached API response is inadequate when an agent can write to a database, issue a refund, search a changing repository, or use a tool that did not exist in the original trace. Raindrop says it creates stateful synthetic copies of the services the agent touches and generates data shaped like the customer's environment.
Frontier model developers have been pursuing a similar testing model internally. In June, OpenAI published research on deployment simulation that used de-identified production conversations to forecast undesirable behavior before model releases. OpenAI reported that realistic deployment simulations gave it a better view of behavior frequency than selected test prompts, while acknowledging that rare failures still require targeted evaluations and red-team testing.
Raindrop is packaging that method for application developers whose agents span models, prompts, tools, and business systems. Simulations is in early access, with Raindrop saying general availability is planned over the coming month.
The founders found the problem while building an agent
Raindrop grew out of the founders' experience building Sidekick, a coding agent and VS Code extension. The team could attract paying users but lacked an effective way to understand why the agent failed after deployment. That problem pushed the founders away from the application and toward the monitoring infrastructure they needed to operate it.
Koticha and Gauba had previously built Opyn, a financial software company acquired by Coinbase. Hylak spent four years at Apple, moving from engineering into the Human Interface design group and working on visionOS. Koticha told Creative Conviction that he and Hylak had lived together, while he and Gauba met at the University of California, Berkeley and built Opyn together.
The trio entered Y Combinator's Winter 2024 batch and positioned Raindrop as monitoring infrastructure for AI agents. The founders' thesis was that conventional application monitoring could identify a crashed service or failed network call, while missing an agent that completed a request incorrectly, forgot an instruction, entered a loop, or took a technically valid action that violated the user's intent.
That framing helped Raindrop secure its $15 million seed less than a year before the Series A. The seed included Figma Ventures, Vercel Ventures, Y Combinator, and founders or executives associated with Replit, Cognition, Framer, Speak, and Notion, according to Raindrop's December 1st, 2025 funding announcement.
A reliability stack takes shape
Raindrop now lists Vercel, Speak, Clay, Framer, AngelList, and Tolan among teams using its products. Raindrop also claims it processes billions of traces each month and serves Fortune 100 customers. Those operating metrics remain company-reported, but the shift toward enterprise buyers explains why Raindrop is expanding beyond incident detection.
Large organizations need evidence that an agent change will behave predictably before it can reach a healthcare workflow, logistics system, payment process, or customer account. Production monitoring finds damage quickly. Simulation offers a chance to block the damaging change.
Raindrop says the new capital will support hiring, particularly in machine learning engineering and go-to-market roles. The hiring plan matches the two problems Simulations must solve: reproducing complex customer environments accurately and convincing enterprises to place sensitive production traces inside another testing system.
CRV is funding Raindrop's attempt to own both sides of agent reliability: the record of what failed in production and the pre-release system meant to prevent a repeat. That loop gives Raindrop a stronger position than a standalone tracing dashboard. Every production failure can become a future test, and every proposed fix can be measured against how users actually interacted with the agent.
The founders are betting that production data will become the most useful evaluation set an agent developer owns. Simulations turns that data from an incident archive into a release gate.