Same backlog, two delivery models: what happened when I let an AI agent build my app

A solo builder’s honest account of migrating a habit tracker off no-code, into an owned stack — and what 3 days of agentic coding actually looked like next to a 10-week sprint plan.

Burndown Comparison AI-assisted vs Agentic Coding

Same backlog, two delivery models

Same 118-point backlog. Two delivery models.

The original sprint plan: 6 sprints, ~10 weeks, June 1 → August 9.

What actually happened: 3 days.

That’s the burndown chart above — built by the same AI agent that did the work, pulled live from the project’s own backlog data.

A few honesty notes, because the chart alone overstates the simplicity:

• This isn’t a human-vs-AI race I ran twice. The “AI-assisted plan” line is the original estimate — what I scoped assuming a developer using AI tools, the normal way most of us plan today.
• The ≈23× is calendar-time compression, not labour-hours. I didn’t work 23x harder. I planned in sprints. The agent didn’t need sprints.
• About 89% of the scope was built autonomously. The other 11% — pricing decisions, product taste, a founding-user idea I killed after arguing myself out of it, and a stack of paperwork — that part stayed entirely human. AI was 0x there. On purpose.

The real unlock wasn’t speed. It was that sprint planning itself assumes a constraint — human coding throughput — that doesn’t fully apply anymore.

I’m still the product owner. I still decide what’s worth building. I just stopped scheduling around a bottleneck that isn’t there.

What would you build if your delivery timeline wasn’t the constraint anymore?

How it started: vibe coding in the summer of 2025

Before the chart above, there was a much smaller story.

In summer 2025 I vibe-coded a habit tracker called Momentum on Base44.

It worked. I used it every day. But I didn’t own it.

This week I migrated the entire app to my own stack — Vercel, Supabase, React — using Claude Code with Fable, Anthropic’s new Mythos-class model.

2 hours. Without copying the entire codebase. Plus 2 new features added along the way.

Now I own every line. Every user. Every data point. The full artefact.

That’s the shift nobody talks about enough — vibe coding gets you to v1 fast, but agentic coding gets you to ownership.

https://momentum-2-change.com

Where this leaves me

Stack, for anyone curious: Vite + React SPA, Supabase (Postgres, Auth, RLS, Edge Functions, pg_cron), Vercel for push-to-deploy, PostHog EU for cookieless analytics, Resend for email, Capacitor for iOS and Android. All tracked in Azure DevOps. Built with Claude Code.

~27 real users on the migrated stack as of writing. An autonomous activation agent runs daily, nudging users who signed up but haven’t checked in — it sent 18 nudges on day one, respecting a 7-day per-user frequency cap. A new user even signed up at 3:38am during an overnight build, which is its own small proof that the thing works while I sleep.

The current bottleneck isn’t code. It’s KVK business registration and a tax ID for Stripe — the unglamorous, very human part of turning a side project into something people can pay for.

That’s the honest state of where this is. More to come as the paperwork catches up with the product.

momentum-2-change.com a Notio Spiritus product

If you’re building something similar — solo, with an AI agent doing the heavy lifting — I’d like to hear where your bottleneck is. Code, or paperwork?