Judgment Labs takes $32M from Lightspeed to fix agents in production

verified

Wednesday, May 13, 2026

The News

Judgment Labs announced $32 million in combined seed and Series A funding, with Lightspeed Venture Partners leading both rounds — doubling down less than six months after its initial investment — joined by Nova Global, SV Angel, Valor Equity Partners, and Dynamic. The pitch: evaluation methods inherited from the chatbot era were built for single-input/single-output assessment, but deep agents produce trajectories — long chains of decisions, search queries, partial results, and self-corrections — and when one fails, the glaring fault is usually buried somewhere in that trajectory, not in the final answer. Customers include teams building deep agents in legal, finance, and customer support. (Business Wire, TechCrunch via Citybiz).

The Read

Lightspeed leading both rounds in under six months is the actual signal, not the dollar amount — that's a firm telling its LPs that agent evaluation is a category where the first credible product becomes the standard, and you pay up to lock the founder in before anyone else priced the second round. The category itself is a tell: a year ago, evals were considered a feature inside LLMOps platforms. They're now a venture-scale company because production agents fail in ways that input-output testing literally cannot see. The harder claim — which Lightspeed is paying for — is that the trajectory data Judgment captures becomes more valuable than the model that produced it, because it's the only place where you learn which kinds of plans actually survive contact with users.

Do This Week

Before your next agent release, write down three failure modes you've personally watched in trajectory logs in the last month — not "hallucination," specific ones (e.g., "agent treats order-level complaint as item-level," "agent rejects refund then closes ticket without escalation"). If you can't name three, you don't yet have the observability you need to deploy an agent in front of paying customers, regardless of what your eval scores say.

For Engineering / founder building agent products