Probe gives research agents a memory of experiments already run

An AI research agent can write the next experiment without knowing why the previous one failed. Probe is building a shared workspace for machine learning research that keeps experiments, compute and accumulated context together, so each new session has more to work with than a fresh prompt.

The startup's launch centers on a familiar research problem: useful information is scattered across tracking tools, GPU infrastructure, papers and the people who remember what happened. Adding agents makes it easier to generate work, but doesn't automatically give that work a coherent history. Probe wants the decisions behind past experiments to travel with the next one.

Its product combines experiment tracking, GPU observability, compute and team visibility with a context layer for agents. Rather than treating a run as an isolated result, the workspace is intended to preserve the surrounding research judgment: what was attempted, which decisions mattered and where the next session should start.

That puts Probe closer to research infrastructure than a chatbot for answering questions about papers. The company describes a stack that also brings literature and datasets into the workflow. Researchers can see what agents are running and steer the work, while agents can draw on shared context instead of repeatedly reconstructing it.

Watch Probe's original launch demonstration.

The distinction matters when the output of an agent is an expensive experiment rather than another block of text. A team needs to know which runs deserve attention, how compute is being used and whether a proposed experiment repeats an approach already tried. Probe presents reduced coordination overhead and less wasted compute as benefits of keeping that information together. Those are company claims, not independently measured outcomes.

Probe has published early benchmarks to make its case. On its website, the company reports a best accepted KernelBench speedup of 50.8 times over eager PyTorch with Probe, compared with 34.9 times for its vanilla setup. It also reports a final InferenceBench speedup of 5.06 times versus 3.49 times in a Mistral 7B Instruct serving experiment using a Modal H100 and shared Git.

Those figures describe specific company-run tests. They shouldn't be read as a promise of equivalent gains across research projects, models or hardware. Their relevance to Probe's pitch is narrower: the startup is testing whether better communication and retained context help agents make more productive use of the same research environment.

Ashwarye Yadav, Richard Wei and Mahit Namburu, three Berkeley EECS founders, are behind the company, which participated in Y Combinator's Spring 2026 batch. Their launch arrives as research workflows increasingly need to accommodate work done by both humans and agents.

Probe's practical proposition is that a research workspace should preserve reasoning as well as results. Experiment logs can show what ran. The harder task is keeping enough of the surrounding decisions intact that the next researcher, whether human or software, can pick up where the work actually stopped.

‍