Raindrop raises $35M to catch AI agents failing silently
Agent infrastructure got the money this week: a reliability layer for agents that fail silently, an accounting agent aimed at construction, and an industrial model in Hefei that now schedules power instead of just answering questions.
Raindrop raised a $35 million Series A led by CRV to catch AI agents failing in production, bringing its total funding to $50 million after a $15 million seed in December. The company reads full agent trajectories — every step a run takes — and flags hallucinated answers, tool misuse and, the part most teams miss, behavior changes introduced when an underlying model is upgraded without any code change on their side. Customers include Vercel, Framer and Clay alongside unnamed Fortune 100 enterprises, and the investor list reads like a technical endorsement: Lightspeed and Y Combinator returned, joined by individual researchers from OpenAI, Anthropic and Thinking Machines. Raindrop also launched Simulations in research preview, which replays a proposed agent change against real production traffic before it ships.
Why it matters: this is the failure mode existing monitoring was never built for. Application performance monitoring watches a process you control; an agent's behavior can shift under you because a vendor swapped a model checkpoint overnight. Raindrop's bet is that agent reliability becomes its own infrastructure layer the way APM did once web apps got too complex to trace by hand. The category already has a price tag — Dynatrace bought Arize for $915 million in August, which we covered in AI Observability: Dynatrace acquires Arize for $915M. The open question is whether the labs bundle this themselves; every one of them already has the traces.
Adaptive raised a $30 million Series B led by Tidemark to sell agentic accounting to construction contractors, taking its total raised to $57 million. Construction finance is a genuinely awkward target: progress billing, change orders, retainage and subcontractor compliance run on spreadsheets and tribal knowledge, and the money is large enough to matter. Adaptive's pitch is agents that close the books and chase the paperwork rather than a chatbot bolted onto a ledger.
The interesting signal is where venture money is going in this cycle — not another horizontal assistant, but narrow verticals with messy, document-heavy workflows that no foundation model solves on its own. Construction accounting is unglamorous and exactly the kind of work where an agent that is 90% right still needs a human to check it, which is why the product's review surface matters more than its model choice.
Lingyang released version 3.5 of its industrial model at the World Manufacturing Convention in Hefei, adding multimodal perception, time-series forecasting and an energy knowledge graph — plus three agents for compute-power coordination, power trading and zero-carbon parks. The company, founded in 2022, is 35% owned by Digital Anhui and 30% by iFlytek, and it is pitching the upgrade as a shift from reactive energy scheduling to model-driven decisions.
This is the Chinese industrial AI pattern in one product: no consumer splash, no benchmark chart, just a version bump aimed at factory and grid economics where a percent of energy saved is real money. Xinhua's preview of the convention frames the three agents as targeting cost reduction and multi-energy coordination, which is the language state industrial policy rewards. Whether the savings are real will show up in deployment data, not the launch.
What to watch: whether Raindrop's Simulations product becomes the standard pre-merge gate for agent changes. If it does, "did this model upgrade change my agent's behavior" becomes a CI check rather than an incident.
If a vendor silently upgrades the model under your agent, whose job is it to notice — yours or theirs? Tell us in the comments.
