Project summary

I will build and publish an open test harness that measures what happens when several AI agents change the same live online store, and whether the combined changes help or hurt. I run Luria, a tool that runs conversion tests on Shopify stores, so I already have real stores, a testing engine and safety controls.

What are this project's goals? How will you achieve them?

Goal: get real world data on conflict between AI agents that act on one business. Today a store has one planner. Soon it may have several agents (copy, offers, ads) that each chase their own metric. I will set up small experiments on stores that opt in. I will compare agents acting alone, taking turns and acting at the same time. The old page always keeps at least 10% of traffic and a kill switch stops any test in 60 seconds. I will check whether changes interfere (for example one agent's discount cancels another's headline win) and whether agents claim credit for the same gain.

How will you know it worked?

I finish 12 weeks of experiments on 2 to 3 stores, publish the harness code on GitHub with anonymized results, and write a short public report. Even a result that says the problem is small in practice counts.

How will this funding be used?

About 120 hours of my time (12 weeks at 10 hours a week) at roughly $50 an hour, so $6,000. The minimum of $3,000 pays for a smaller version with one store and fewer experiments. Hosting and test tooling I already pay for through Luria.

Who is on your team? What's your track record on similar projects?

Just me. I am a UC Berkeley CS student (Data Science minor, class of 2027). I built and sold two companies: SAT Masterclass grew from $50k to $1.2M ARR and Stutoring was sold at $900k ARR. Since Feb 2026 I have built Luria, which has multi-arm tests, a held-out control, heatmaps, session replay and an AI planner, and three real stores use it. A change only goes live at 95% certainty after 200+ visits per version. Code: https://github.com/baneets. Site: https://www.luriart.com. I have no published papers and no formal AI safety research experience.

What are the most likely causes and outcomes if this project fails?

Stores may not have enough traffic to show interference, so the results are noisy. Then I will report that clearly and share the harness anyway. I could also run out of time next to school and Luria, so I would cut the scope instead of cutting safety.

How much money have you raised in the last 12 months, and from where?

$0. I have not raised outside money. Luria earns about $1,000 a month from subscriptions.