← Back to feed

Judgment Labs announced $32 million in combined seed and Series A funding on May 12, targeting the gap between how AI agents fail in production and what standard input/output evaluations actually catch.

Lightspeed Venture Partners led both rounds. The return in under six months is notable: Lightspeed partner James Alcorn said the firm “led the seed because the bet was obvious, and we led the Series A because the results have been extraordinary.” Additional investors include Nova Global, SV Angel, Valor Equity Partners, and Dynamic. Angels include Stanford Professor Chris Manning and the founders of DoorDash and Mercor.

The Problem

The core argument is structural. Standard evals measure inputs and outputs — they miss the trajectory in between. A deep agent running a multi-step task produces a long chain of decisions, tool calls, search queries, and memory operations. Any one step can be where things went wrong. Input/output evals see only the beginning and end.

Judgment’s platform evaluates the full trajectory. It tracks what an agent decided, what it searched for, what tools it called, and what it remembered — then surfaces recurring failure patterns and turns production data into training signal for improvement.

The company ships four research areas currently productionised:

• Agent Search: Query across agent trajectories at a behavioural level, beyond keyword matching on inputs and outputs.
• Agent Judge: Trajectory-level evaluators using harnesses to deliver cheaper and more accurate scoring than LLM-as-a-judge on raw outputs.
• Behavior Discovery: Surface failure modes and usage patterns from unlabeled production runs.
• AutoRubrics: Automatically construct and refine evaluation rubrics from verifiable signals in production data.

The open-source SDK (judgeval) integrates with OpenAI, Anthropic, Google GenAI, LangGraph, and the Claude Agent SDK. Evaluation runs are asynchronous and server-side, adding no latency to the production agent.

The Team

Three co-founders, all in their early twenties and childhood friends: Alex Shan (CEO), Andrew Li (Chief Scientist), and Joseph Camyre (CTO). The company is headquartered in San Francisco and describes itself as an applied-research lab that productionises research findings.

What the Capital Is For

The majority of the $32M goes toward hiring AI researchers and engineers in San Francisco. A secondary allocation expands the forward-deployed engineering team that works on-site with customers — a customer-success model common in developer-tooling companies where adoption depth matters more than breadth.

Why It Matters

The agent evaluation market is crowded at the top — Braintrust, LangSmith, and others offer observability and evals for LLM apps. Judgment is positioning at the harder end: long-horizon agent trajectories where the number of intermediate steps can run into the thousands and where “did the output look right” is genuinely insufficient.

As agentic workloads scale — every major lab shipped an agent SDK or managed agent product in the last 90 days — the tools that surface why an agent failed at step 47 of 200 become load-bearing infrastructure. Lightspeed’s rapid return suggests at least one top-tier firm believes that market is real now, not in two years.

Key Numbers

• Funding: $32M combined seed + Series A
• Lead investor: Lightspeed Venture Partners (both rounds, <6 months apart)
• Platform: Already in production at agent-native companies
• Open source: judgmentlabs/judgeval on GitHub, integrates Claude Agent SDK, LangGraph, OpenAI, Anthropic, Google GenAI