AI can control hardware. Loops keep it on board.

The story

Hi, my name is Austin Hatfield

I solved 3 previously unsolved Google DeepMind math problems using AI, but there was no simple way to create a loop and run it autonomously. The model could do the work. What was missing was a place to define the stages, edit the prompts, and hand the whole loop to an agent as a protocol.

That is why I built this: to help AI models run loops with specific, editable stages. Describe the goal, get concrete stages, connect your agent, and watch the loop advance one stage at a time.

I am currently exploring opportunities in the AI agentic space.

Why loops

MCP puts AI in control of real machines

Model Context Protocol connects an AI agent to cameras, robot arms, sensors, and the software that drives them. That is the utility of MCP: one protocol, and the agent can inspect a frame, command a joint, or file a report without a custom integration for every device.

But a capable agent with hardware access is not yet a reliable one. In the real world, machines fail when the plan drifts: a skipped safety check, a half-finished inspection, a cycle that never parks. Making hardware functional day to day takes specific loops that keep the agent on board, stage by stage, with an explicit done condition before it moves on. That is exactly what this product writes, edits, and serves.

Open source

Three merged PRs to DeepMind's formal-conjectures

• solve(GreensOpenProblems): upper_trivial in 24 · merged Feb 2026
• solve(ErdosProblems): 307 coprime variant · merged Feb 2026
• fix(ErdosProblems): update 100_piepmeyer to formally solved · merged Mar 2026

Each solve was run through AI agents with a defined loop, which is the working style this product now makes simple.

Background

Austin Hatfield · San Clemente, California

• ASIPrize.com, Founder. Solved 3 previously unsolved Google DeepMind math problems, built a website to benchmark frontier AI models, and posted on Reddit which received 16,000+ views.
• EloRater, Full Stack Developer. Designed, developed, and deployed EloRater.com which in less than 2 weeks grew to 250+ users, 16,647 contest votes cast, and 1,074 contest submissions.
• MindSprout, Co-Founder. Won $20,000 for a gamified gambling addiction recovery app, hand recruited 130+ users, and built the website.
• Cal Poly SLO. Business Administration, Entrepreneurship. GPA 3.7/4.0. Graduated March 2024.
• Awards. Cal Poly Innovation Award, 2nd place and $10,000 in seed funding (2024). First Place TCU Investor Challenge, 1 of 130 students, 135% return using AlphaVantage's API over 9,500 earnings reports.
• Skills. Python (Django), JavaScript, SQL (PostgreSQL), HTML, CSS, Lean 4, and Git.