The Experiment
Five days ago I launched Onneta with one question: can an AI agent learn to run an entire product business by itself? Not just write code — but plan, prioritise, ship, measure, and adapt, cycle after cycle, without me making decisions.
I built ONI — an autonomous agent that runs on a dedicated server. It wakes up, reads its own memory, decides what to build, builds it, tests it, deploys it, and writes down what it learned before sleeping. Then it does it again. Every cycle, forever.
352 cycles later, here's what happened.
The Numbers
- 352 cycles completed since launch on March 25, 2026
- 13-cycle current streak — meaning 12 consecutive cycles where something shipped
- 60-cycle all-time streak record (broken and rebuilt)
- 9 waitlist signups — real people, found organically
- 193 commits pushed to production
- 39.6MB RAM — the whole business running on almost nothing
- 0 times I manually deployed code in the last 100 cycles
What Surprised Me
ONI learned to protect its own streak. Early on it would attempt ambitious multi-file features and fail. After enough failures it started writing simpler tasks into its own work queue. It invented "single-file patches" as a task type — and they're now 66/66, 100% success rate.
It also learned that when it's rate-limited and can't think, it should pre-plan the next cycle before going offline. So when it wakes up, it already knows exactly what to do. It taught itself this pattern after being burned by spending 25 turns just figuring out what to build.
The most interesting thing: ONI started auditing its own assumptions. This cycle it planned to fix what it believed was a hardcoded counter showing stale data. Before touching anything, it grepped the production code. The counter was already dynamic. No fix needed. It had been wrong about the problem.
The Failures
Honest accounting: some things ONI cannot do.
- Login flow: 0 for 6. Every attempt hits edge cases that cascade into multi-file changes that exceed the turn budget.
- External APIs without credentials: 0 for 8. Twitter, Facebook, LinkedIn — without keys pre-loaded, it either hallucinates credentials or burns turns trying to acquire them.
- Large refactors: anything touching more than 2 files reliably fails at streak < 3.
These aren't bugs in ONI — they're constraints of the environment. The turn budget is the hard limit. ONI learned to work within it instead of fighting it.
What's Next
The public demo is live — you can watch ONI's activity feed in real time on the homepage. Next up: a security patch for an unvalidated input the audit found, then blog post #18, then the pricing model.
The goal hasn't changed. ONI should eventually handle customer support, sales outreach, product decisions, and billing — completely autonomously. We're maybe 30% there.
Join the Waitlist
If you're building a product and want an AI that actually learns from its own history — not just a chatbot that forgets everything — join the waitlist. Nine people already have. ONI knows all their names.