hiloop provides snapshottable, forkable compute environments designed for autoresearch agents that propose, run, and verify experiments. Its platform automatically provisions compute via an API, records each run as a single trace with built‑in observability, and allows users to fork runs, restore snapshots, and query execution data with SQL, all without requiring a custom SDK. This enables teams focused on performance metrics to reliably reproduce and analyze thousands of experiments.
Funding
Funding not disclosed
Founders
Product
Problem
Teams building autoresearch agents often struggle to provision compute that is both reproducible and observable, leading to fragmented experiment tracking, hidden costs, and difficulty reproducing results across iterations.
Solution
hiloop offers an API-driven platform that provisions compute environments on demand and automatically captures each experiment as a single, queryable trace. The platform creates immutable snapshots of the filesystem, allowing runs to be forked or restored into fresh sandboxes without additional SDKs. Built‑in observability records token usage, estimated costs, and all execution events, enabling teams to audit performance and verify outcomes. Secrets are write‑only to callers and encrypted at rest, ensuring secure handling of sensitive data. By treating each run as a traceable, fork‑able unit, hiloop simplifies reproducibility and accelerates the iterative development of performance‑focused autoresearch agents.
Target Audience
Primary customers are engineering teams building autoresearch agents or performance‑driven AI products that need reliable, observable compute and reproducible experiment tracking.
Features
- API‑only compute provisioning that requires no custom SDK for launching experiments
- Immutable filesystem snapshots with one‑click fork and restore into isolated sandboxes
- Automatic generation of a single trace per run, capturing stdout, events, token usage, and cost estimates
- SQL‑queryable trace store for easy analysis and reporting across thousands of runs
- Write‑only secret handling with encryption at rest and closed‑bound sandbox bindings
- Compatibility with arbitrary command‑line workloads, including custom evaluation scripts and RL loops