ARIA: Sustained Autonomous Research

The problem

The One-Shot Ceiling

Nearly every autonomous research system shares one shape: an agent loop that runs, produces an artifact, and exits. That works for a single experiment. It cannot accumulate signal over weeks, react to a critic's verdict on yesterday's result, or stop itself when a research direction collapses. Real science is sustained.

• Agent loops that produce one artifact and terminate
• No memory of which directions stabilized and which collapsed
• Failures halt the run instead of triggering recovery
• Cloud GPU costs make months-long autonomous operation impractical

What I built

A Persistent, Self-Healing Flywheel

ARIA is a virtual research collaborator built to run for months. A weighted pool of research ideas competes on a single scoring function, the highest-weighted idea runs locally first and scales to cloud only if it earns it, a critic on a different model family posts a verdict, and failures classify themselves and re-enter the pool as recovery work.

• Six-stage flywheel: Pool, Score, Execute, Analyze, Critique, Self-heal
• Human-inserted and auto-generated ideas compete on the same scoring function
• Cost-asymmetric hybrid execution: local first, preemptible cloud only when earned
• DEBUG-driven recovery: 97.8% autonomous resumption across 321 failure events
• One portable kernel, re-pointed across three organizations and their domains

Timeline

How it went

1. Dec 2025ARIA deployed on the homelab DGX: scored idea pool + six-stage flywheel
2. Dec 2025 - Apr 202618-week run: 7 instances, 4 domains, 19,364 commits, 97.8% resumption
3. Mar 2026Kernel re-pointed inside a pharma R&D org, on public data, as an adaptation test
4. May 2026Provisional patent filed (patent pending)
5. Jun 2026ARIA-SE forked onto a borrowed NVIDIA 8xH100 grant node
6. Jul 2026Paper published on Zenodo; kernel released under Apache 2.0; ARIA-SE at 8,700+ sessions

What it taught me

Lessons

1. Sustained beats one-shot: the value is in signal that accumulates over weeks
2. A single scoring function lets human and machine ideas compete on merit
3. A critic on a different model family catches what a model grading its own work never would
4. Self-healing recovery is what keeps an unattended system alive across hundreds of failures
5. Build the kernel once, re-point it at a new domain: portability was the thesis, and it held across three organizations
6. Navari is agents working for a scientist on demand. ARIA is agents doing research unattended. Same builder, opposite shape, and each taught me something about the other.