Raindrop raises $35M to test agents against their production past

Raindrop has raised a $35 million Series A led by CRV, bringing its total funding to $50 million, while opening early access to a product that tests AI agent changes against traces of how those agents behaved in production. The startup says Simulations can run on every pull request and flag unexpected behavior before an update reaches users.

The product moves Raindrop one step earlier in the failure cycle. Its existing software monitors deployed agents for silent failures. Simulations takes production traffic and known test cases, reruns them against a proposed change, and applies anomaly detection to identify behavior that moved in an unexpected direction.

Tests usually cover failures someone imagined

Traditional evaluations begin with examples and expected outcomes selected by a team. That is useful for checking known risks, but it is weaker when a model, prompt or tool change creates a failure no one thought to encode as a test.

Raindrop says its approach uses a company’s own production history as part of the evaluation set. A changed agent harness is exposed to previous traffic and existing tests, and its new trajectories are compared for anomalies. The aim is to answer a practical release question: what else changed when engineers changed one piece of the agent?

The distinction matters because agent failures do not always produce a software error. An agent can complete every tool call successfully while misunderstanding an instruction, choosing the wrong sequence of actions or producing a plausible answer that violates the user’s intent. Those are behavioral regressions rather than conventional crashes.

Replaying calls is not enough

An agent’s environment can include databases, payment systems, messaging tools and APIs whose state changes after every action. Cached tool responses cannot represent a newly added tool or a workflow that follows a different path from the original trace.

Raindrop says Simulations reconstructs the world surrounding the agent in a harness-agnostic way. The company compares each observed tool call to a hole in Swiss cheese: a partial view of the underlying state that can be used to generate a simulated environment. That claim is central to the product, but Raindrop has not published independent measurements showing how accurately the simulated worlds predict failures after deployment.

Production history becomes a release gate

The company says it developed Simulations with Fortune 100 partners. Early access is now expanding, with general availability planned over the next month. Researchers and executives associated with Anthropic, OpenAI and Thinking Machines also participated in the financing, according to Raindrop.

The unusual asset here is not another benchmark. It is the archive of everything an agent previously did in front of real users. Raindrop is betting that this history can become a regression suite, allowing every production failure to become evidence against the next bad release. The concrete unknown is whether its generated environments remain faithful when agents encounter tools and situations that never appeared in the original traces.