Reific trains AI agents to reason as live data changes

Reific is building training environments for a problem that does not wait for an AI model to finish thinking: data that keeps changing. The Fall 2026 Y Combinator startup develops reinforcement-learning environments in which signals arrive on a clock and an agent's actions are evaluated against subsequent outcomes.

The company is targeting the frontier labs that train agents. Its premise is that competence on text and coding tasks does not automatically translate into useful decisions about time-series data. A model may have to interpret multiple changing signals, decide when to act and account for information that arrives while it is already working.

Reific groups its environments into three task families: personal agents, industry, and science and compute. Its public company profile says the environments draw on real, de-identified records. The launch describes examples ranging from heartbeats to bioreactors and GPU clusters, illustrating the types of signals it wants models to learn from rather than announcing deployments in those settings.

Watch Reific's original launch video.

The distinction is in the feedback. Rather than reward an agent solely for reproducing an expert's answer, Reific says its environments grade actions by what actually happens. That makes the timing and consequences of a decision part of the training problem. The company's ambition is to move beyond matching human performance, though the results it has disclosed remain company-reported benchmarks.

In its launch post, Reific says the best frontier model averages below 50% in each of its three task families. It also reports that training an open model in its environments lifted the full-solve rate on unseen tasks from 25.6% to 39.5%, with no loss on general benchmarks. That is a 13.9-percentage-point increase, or roughly 54% relative improvement. The launch does not identify the model or provide enough evaluation detail to independently assess that result.

Founder and CEO Antanas Žilinskas brings a background in time-series research and engineering. His public YC profile describes him as first author of two ICLR 2026 papers on rare-event forecasting and a former co-founder and CTO of Entangl, a YC Summer 2024 company. He also says he worked on superconductors at CERN and helped take engineering AI into enterprise deployments.

Reific's proposed role is training infrastructure, not a finished application for patients, factories or data-center operators. For prospective users, the relevant question is whether improvements inside its environments transfer to the signals and decisions their agents will face outside them. The public launch gives an initial performance claim and a defined training approach, but does not announce pricing, named customers or production results.