Raindrop Raises $35M for AI Agent Simulations

Production is where an AI agent learns the difference between a clean test and messy customer reality. Raindrop wants that lesson to survive long enough to challenge the next pull request. The San Francisco company has raised a $35M Series A to connect what fails after deployment with what should be blocked before release. CRV led the September 17, 2026 financing, with existing investors Lightspeed Venture Partners and Y Combinator participating alongside researchers and executives from Anthropic, OpenAI and Thinking Machines. The round brings Raindrop's total announced funding to $50M after a $15M seed.

The financing arrived with Raindrop Simulations, a research-preview product that replays production traffic and existing tests against proposed agent changes. The strategic move is larger than another evaluation dashboard: Raindrop is trying to turn production failures into reusable release evidence, connecting what broke yesterday with what should be blocked tomorrow.

What Raindrop Raised and Who Backed It

The company announcement describes a Series A led by CRV and $50M in total funding. Axios Pro identified the new round as $35M, which aligns with Raindrop's previously announced $15M seed led by Lightspeed Venture Partners in December 2025.

CRV is the new lead investor. Lightspeed and Y Combinator returned, while Raindrop also named researchers and executives at Anthropic, OpenAI and Thinking Machines as participants without identifying the individuals. The company did not disclose its valuation, security terms, ownership changes, board changes, check sizes or investor allocations.

Raindrop said the capital will accelerate anomaly-detection research, expand enterprise adoption and advance Simulations. Its current hiring plan spans machine learning, infrastructure, product, security, sales and marketing, matching the technical and commercial work required to move from a developer tool into enterprise release infrastructure.

From Production Alarm to Release Gate

Raindrop began with a production problem. AI agents can return a plausible answer, call the wrong tool, enter a loop or shift behavior after a model change without triggering the error signals conventional software monitoring expects. Raindrop reads agent trajectories and looks for semantic anomalies across the full interaction rather than waiting for a service to crash.

That monitoring layer helps a team discover what already went wrong. Simulations moves the evidence earlier in the lifecycle. Raindrop says the product runs on every pull request, replays production traffic and existing tests against the proposed agent change, then applies anomaly detection to find expected and unexpected behavior differences.

The difficult part is recreating the world surrounding the agent. A cached response cannot fully represent a tool that writes to a database, issues a refund, queries a changing repository or did not exist in the original trace. Raindrop says its approach simulates the tools and state around the agent so teams can evaluate a new harness rather than replaying a frozen transcript.

Why Production-Derived Testing Matters

Traditional evaluations remain useful for known behaviors and targeted risks. Their limitation is inventory: a team must know enough about a failure to write the test. Production traffic contains the awkward combinations of user intent, tool state and model behavior that no fixture anticipated.

OpenAI's deployment-simulation research offers independent support for the broader method. OpenAI found that realistic conversation contexts and carefully simulated tools can improve pre-release estimates of model behavior in agentic settings. The research also keeps the boundary clear: deployment simulation complements targeted evaluations, red-teaming and tail-risk analysis rather than replacing them.

That distinction matters as agent tasks become longer and more consequential. METR reports that the time horizon of tasks frontier agents can complete at 50% reliability has roughly doubled every seven months since 2019. More autonomy creates more opportunity for useful work, but it also gives a small behavioral change more room to compound before a person notices.

The Founders Are Rebuilding Their Own Feedback Loop

Raindrop was founded by Zubin Koticha, Ben Hylak and Alexis Gauba. Y Combinator lists Koticha as Founder/CEO and the company as a Winter 2024 participant founded in 2023. Hylak is Raindrop's co-founder and CTO, while Gauba is a co-founder.

Koticha and Gauba previously built Opyn, a financial-software company acquired by Coinbase. Hylak spent four years at Apple and worked on the Human Interface team behind visionOS. The three founders moved toward agent monitoring after encountering the difficulty of understanding silent failures in an AI product they had built, making the feedback loop more than a market thesis borrowed from a slide.

Raindrop says Vercel, Framer, Clay, Speak and Fortune 100 enterprises use its products. Its careers page reports more than 200 customers and a team of fewer than 20. Those are company-reported operating signals, not audited measures of revenue, retention, detection accuracy or simulation fidelity.

What the $35M Now Has to Prove

Raindrop is entering a crowded field that includes evaluation platforms, observability vendors, tracing tools and internal testing systems. Its sharper position is the connection between production monitoring and pre-release simulation: every discovered failure can become part of the next release decision, while each proposed fix can be tested against behavior drawn from actual use.

The promise will be judged by fidelity and trust. Simulations must reproduce enough of a customer's changing environment to catch meaningful regressions, control false alarms, protect sensitive production data and fit inside engineering workflows that already have plenty of gates. Enterprise buyers will also want evidence that the system improves release quality without turning every model or prompt update into an investigation.

The Series A gives Raindrop more room to build that evidence while its customers place agents inside healthcare, logistics, finance and other high-stakes workflows. The enduring asset may become the history each team accumulates. Every incident adds another record of how its agents fail, how the surrounding systems respond and which changes deserve to reach the people depending on them.