Raindrop

What Raindrop is building

Raindrop is Sentry for AI agents, and the core idea is that agent failure is a detection problem rather than a testing problem. Agents run for hours, call thousands of tools and fail convincingly: a hallucinated answer, a wrong tool call, a behaviour shift after a model upgrade. None of that throws an exception, so nothing in a conventional observability stack fires. Raindrop reads production agent trajectories and applies anomaly detection to find the semantic failures, then clusters them across runs into Issues with confidence scores, the way a crash reporter groups stack traces. The primitives on top of that: Tracing (every message, tool call and decision as a span tree), Signals (custom classifiers for behaviours you care about), Experiments (feature-flag A/B tests judged against baseline production traffic), a Triage Agent that investigates a failure and pushes follow-ups into Slack, the web app or MCP, and Workshop, an open-source local replay and debugging tool with 1.1K GitHub stars. Pricing is public: free hobby tier at 1,000 events a month, Pro at $299 a month plus $0.003 per event, Enterprise with SSO, audit logs and edge PII redaction. Self-hosting is in beta; SOC 2 Type II is done. The new piece, launched with this round as a research preview, is Simulations. It takes real production traffic and existing test cases, replays them against a proposed agent change inside stateful synthetic copies of the services the agent touches (databases, comms tools, payments, APIs), then runs the same anomaly detection over the resulting trajectories and reports regressions, cost changes, output drift and tool errors before the PR merges. That closes the loop: production detection tells you what broke, Simulations tells you whether the fix breaks something else. Customers named publicly: Vercel, Framer, Clay, Speak, AngelList, Browserbase, Tolan, Howie, plus Fortune 100 enterprises in healthcare and logistics. Tolan's case study puts a 27.8% reduction in memory issues on the product.

Why this matters

The pitch is in Zubin Koticha's own sentence: 'When an agent fails, it does the wrong thing convincingly at scale until someone happens to notice.' Traditional monitoring assumes failures are loud. Agents make them quiet, and the surface area grows with every hour an agent runs and every tool it can call. That is a new category of production risk with no incumbent, which is why a security-style detection company can exist here. CRV's Reid Christian frames it the same way: 'Raindrop treats agent failure as a detection problem, the way a security company would.' The team is built for that thesis, with fraud-detection engineers from Robinhood, anomaly-detection people from Square and security engineers from Segment, Semgrep and Socket.dev. And the customer list (Vercel, Clay, Framer, AngelList) is the set of companies shipping agents to the most users right now, so the failure data Raindrop sees is ahead of everyone else's.

Investors: CRV (Reid Christian), Lightspeed Venture Partners, Y Combinator, Angels: lead researchers from OpenAI, Anthropic and Thinking Machines