Research must flow.

Compute, experiment state, and observability for agents that do ML research. hiloop is the infrastructure under the loop.

Talk to the founders
HILOOP · INFRASTRUCTURE FOR AUTONOMOUS RESEARCH

Everything an agent needs to run ML research.

WHAT WE'RE BUILDING

The best researchers spend their weeks provisioning hardware, untangling broken runs, and hunting for a result from three weeks ago. That work should belong to agents — but the infrastructure underneath was built for humans. We're building it for agents.

Compute

Hardware an agent can provision itself, in seconds. Snapshot, branch, and roll back are first-class operations — an experiment behaves like a git branch, cheap to try and safe to abandon.

SNAPSHOT · BRANCH · ROLLBACK

Any harness

We're the layer underneath, not another harness. Bring a homegrown eval script or an off-the-shelf RL loop — the primitives are the same either way.

YOUR LOOP · OUR SUBSTRATE

Memory

Every run leaves a trace: code, data, logs, results, and the reasoning behind them. The record of what's been tried — including every failure — is what the next agent learns from.

EVERY RUN → ONE TRACE

Hill climbing, in a loop.

THE LOOP
01 Provision The agent asks for hardware and gets it in seconds. No tickets, no waiting on a human with cluster access.
02 Run The experiment runs with everything recorded — code, data, logs, and the agent's reasoning, in one place.
03 Branch Snapshot the run, fork it, try a small change. If it fails, roll back and try another way.
04 Learn What worked joins the record. The next climb starts higher than the last one ended.
The bottleneck isn't GPUs. It's experiment state.

Every lab rebuilds the same plumbing from scratch — state in one place, artifacts in another, the reasoning nowhere at all. We think the record of every experiment, every branch, every honest failure is the most valuable thing a lab owns. So that's what we're building: a place where it all flows into one stream, and stays.

HILOOP — THE BET · 2026