The Pulse
Raindrop Announces Series A, Says It Has Raised $50M in Total Funding
Raindrop says it has raised $50 million in total funding, led by CRV, as the startup expands from monitoring AI agents in production to testing their behavior before new code reaches users.
AI.info Team ·
Raindrop says it has raised $50 million in total funding, led by CRV, as the startup expands from monitoring AI agents in production to testing their behavior before new code reaches users.
The company announced the funding on September 17, 2026, alongside Raindrop Simulations, a product that runs on every pull request and shows engineers how a proposed change could alter an agent’s behavior. The investor group includes researchers and executives from Anthropic, OpenAI and Thinking Machines, according to Raindrop’s announcement.
From production alerts to pre-release testing
Raindrop built its business around failures that conventional software monitoring can miss. The company says its work over the past year has focused on unsupervised, real-time detection of problems in production systems, from bugs that annoy users to issues affecting health and safety.
Raindrop says its customers now include large enterprises in healthcare, logistics and other sectors, alongside the startups that formed its original customer base.
Simulations extend that monitoring system into the development process. Rather than relying only on fixed test cases, the product replays real production traffic and existing tests against a proposed change to an agent harness. Raindrop then applies its anomaly-detection system to the results, allowing teams to detect unexpected behavior changes before they reach production.
Why ordinary evals miss agent failures
Traditional evaluations generally measure behavior against scenarios engineers have already written down. That approach can identify known failure modes, but it may not catch unexpected changes in an agent’s behavior.
Raindrop says its simulations are designed to show “what their change will change” before deployment. The system lets AI engineers measure performance against their own test cases while also looking for unexpected behavior changes.
You have to actually simulate the entire world around the agent.— Zubin Koticha, Ben Hylak and Alexis Gauba, Raindrop announcement authors
Raindrop says the technical challenge is making simulations work across different agent harnesses. Replaying network requests or using cached tool responses is not enough when an agent adds a tool that never appeared in the original trace, or when a team moves its system to a different harness.
Raindrop’s bet on simulated worlds
Raindrop describes each tool call as a view into the state of the world surrounding an agent. Its simulation system uses those observations to create simulations of the environment around the agent.
The approach draws on methods already being developed by frontier AI laboratories. Raindrop points to OpenAI research on deployment simulation, which uses de-identified production conversations to estimate how a candidate model may behave, and to Anthropic’s work on synthetic environments for training and stress-testing agents.
Raindrop says it has been developing Simulations with Fortune 100 partners. The company is now expanding early access and plans to make the product generally available over the coming month.
A larger market for watching autonomous software
The funding gives Raindrop more room to pursue a problem that becomes harder as agents take longer action chains and gain access to business systems. Monitoring a final answer is not enough when a failure may begin several tool calls earlier or emerge only after an agent interacts with external data.
Raindrop says it is hiring across the company, with a focus on go-to-market and machine-learning engineering roles. The immediate product milestone is narrower: early access is expanding now, with general availability planned within the next month.