Problem

AI teams can often get general-purpose models to perform adequately, but pushing them further on a measurable task requires a large, expensive loop of proposing changes, running experiments, comparing results, and reproducing the winners. The bottleneck is not merely model intelligence: it is the engineering infrastructure needed to provision compute, preserve state, track thousands of runs, control costs, and verify that an apparent improvement is real. Hiloop’s killer use case is therefore automated model and agent improvement, especially when performance depends on exploring many training recipes, prompts, tools, harnesses, or systems configurations.

The economic pain is the cost of GPU time and scarce research-engineering attention. Hiloop’s own benchmark campaign illustrates the opportunity: stock coding agents ran 4,188 experiments in two days across 50 B200 GPUs, with the system recording and validating the search rather than relying on a single lucky run.

Product / Service

Hiloop is infrastructure for “autoresearch,” in which agents propose, execute, and verify experiments against a chosen metric. Teams provide a task, an existing agent or model, and an evaluation; Hiloop orchestrates the search across training, data, supervised fine-tuning, post-training, continual learning, prompts, tools, and systems, then returns the best verified improvement. The platform can run hosted by Hiloop or in the customer’s cloud.

The product combines API- and CLI-accessible compute with snapshotting, forking, and restoring of experiment environments; built-in observability; persistent experiment lineage; and statistical verification. Any agent harness can be used, while every run is captured as a trace with model tokens and estimated cost queryable through SQL. The benefit is a reproducible, inspectable research loop that lets teams scale experimentation without building a bespoke orchestration, sandboxing, and evaluation stack.

Market

Hiloop sits at the intersection of AI developer tools, machine-learning infrastructure, agent evaluation, and reinforcement-learning or post-training infrastructure. Its distinctive positioning is broader than experiment tracking or evaluation alone: it aims to operate the full recursive-improvement loop, from compute and state management through automated search and verified model or agent gains. Adjacent competitors include LangSmith for agent testing and evaluation, Weights & Biases for experiment tracking, and Modal for elastic GPU infrastructure; these overlap with pieces of Hiloop’s workflow but are not presented in the available evidence as direct equivalents.

The company is an active San Francisco startup founded in 2026 and listed in Y Combinator’s Summer 2026 batch, with a two-person team and an early-access invitation rather than publicly disclosed customer or revenue metrics. Its strongest public traction is technical: Hiloop reports a 0.9016 median validation score on Karpathy’s autoresearch benchmark after 4,188 experiments, versus a cited prior B200 result of 0.9109. That demonstrates product capability and creates a credible wedge into agent and model training, but the public evidence indicates an early commercial stage, not an established revenue business.