NeoSigma builds infrastructure for teams running agents in production, turning live sessions into private evaluation environments they can replay before a change ships. The launch product, NeoEval, reads traces, failed sessions, and existing verifiers, groups similar failures, and builds a sandbox eval for each kind of failure. Each eval runs the real agent at a chosen commit, with the original task, a simulated user, and the team's own checks, so a prompt, model, or tool change can be tested for regressions. On one published customer week, those evals covered 42 of 51 agent failures from the prior two weeks. Frontier models keep improving on general tasks, while agents in production still fail on a company's own data, tools, and edge cases, leaving work incomplete until a person steps in. NeoSigma captures those failures, along with retries, corrections, escalations, and overrides, and turns them into evaluation environments the organization owns. The environments are meant to move with users and workflows, and to be where teams optimize for higher task success, lower cost, and faster runs. Gauri Gupta, co-founder and CEO, was early at Parallel, where she built deep research agent infrastructure and led evals, after foundation model research at Apple and MIT. Ritvik Kapila, co-founder and CTO, previously led foundation model training, data, and evals at Essential AI. Based in San Francisco, the company is already working with AI-native companies and large enterprises, and is opening access through a waitlist. Its site lists backing from Jeff Dean, Karthik Narasimhan, Anmol Gulati, and angels from OpenAI, Google DeepMind, Cognition, World Labs, and Mercor.