We Rebuilt Our Agency With an AI Agent: 34 PRs in 11 Days
Between August 7 and 17 we repositioned NavyaAI, rebuilt the site, rewrote the pricing model twice, and shipped two new service lines — 34 merged pull requests, every one public, most of the code written by an AI agent under human gates. Here is exactly how it worked, including where the agent was wrong.
Between August 7 and August 17, we repositioned NavyaAI, rebuilt the website, rewrote the pricing model twice, shipped four new service pages, and replaced the entire visual system. Thirty-four pull requests, merged to a public repository, in eleven days.
Most of the code and copy was written by an AI agent. None of the decisions were.
We sell exactly this — production agentic delivery with human gates — so the rebuild was also the demo. Here is how it actually worked, receipts included.
The trigger
Our lead flow had collapsed. The diagnosis, run against our own database and Search Console, was blunt: the site was positioned for one audience (teams optimizing $20K+/month AI bills) while the business needed a second one (teams that want AI built). Fixing that meant repositioning, repricing, and rebuilding — the kind of project agencies quote in quarters.
We gave it eleven days and an agent.
The operating model: agent executes, human gates
Every unit of work followed the same loop:
- Plan first. The agent proposed a written plan — files, approach, verification steps. Nothing was built until the plan was approved, and plans were rejected and redirected often.
- Build with proof. Code had to compile, the full site had to build clean, and every merge deployed to production — where the agent polled the live site and verified its own changes against real markers, then captured full-page screenshots for review.
- Human taste, human truth. I approved or rejected designs from screenshots. I set every price. I struck every claim we could not stand behind. The agent drafted; I decided.
That loop is why 34 PRs in 11 days did not produce a mess. The speed came from the agent. The safety came from the gates.
What eleven days actually contained
A compressed timeline from the public PR history:
| Days | What shipped |
|---|---|
| Aug 7 | Lead-recovery sprint: instant-unlock report gate, benchmark report, SERP fixes |
| Aug 9 | Campaign pages, Cal.com booking attribution + webhook, homepage redesign, services architecture — eleven PRs in one day |
| Aug 10 | Team pricing built, then rebuilt from scratch the same day when the first version failed the founder's smell test; positioning reconciled against an external expert audit; services redesigned — twelve PRs |
| Aug 12 | Gradient-merged header, ambient neural canvas, scroll-aware navigation |
| Aug 17 | Agentic AI service page, generated art wired, keyword harvest: two more service pages |
The Aug 10 pricing story is worth the detail. The first version priced engineers at US-adjacent rates, benchmarked properly against Toptal-class platforms. It built cleanly, deployed cleanly — and looked fake, because buyers know Indian agencies exist. Judgment call, human made: we threw it away the same day and rebuilt pricing anchored to our real rupee compensation bands, published openly. The agent executed both versions flawlessly. Only one of them was right, and no eval catches that. A founder does.
Where the agent was wrong
Publishing failures is house style, so: the agent shipped currency approximations that were 5–6% stale (its EUR/GBP/AUD rates lagged real August 2026 FX until a verification pass against live data corrected them). It calculated our site header as 60 pixels tall when the real height was 65, leaving a five-pixel white band at the top of the page that I caught from a screenshot. It once started editing the wrong section of a file and had to be reverted mid-flight.
None of these reached users for long, because nothing the agent produced was assumed correct. Builds gated merges; live verification gated "done"; screenshots gated design. An agent without those gates would have shipped all three mistakes silently. That is the whole thesis: the demo is the easy 80% — the gates are the production 20%.
What stayed human, permanently
Worth being precise about, because "AI built our site" headlines usually hide this list:
- Positioning. "Build AI that survives production" came out of an expert audit and founder judgment, not a model.
- Pricing truth. Every number on the pricing pages is a real compensation band plus a flat delivery layer. The agent enforces the math (prices derive from one source file); a human chose the model.
- Claims. Every public statistic on the site is published with raw data or was struck. The agent was explicitly forbidden from inventing numbers, testimonials, or clients — and the one time a client was misfiled as a product, a human caught it.
- Taste. Screenshots were approved or rejected in plain language ("this section looks good", "this seems off") and the agent iterated until the human said ship.
Autonomy without control becomes risk. Intelligence without governance becomes liability. We put that principle on our homepage before we applied it to ourselves; the rebuild is what it looks like in practice.
The uncomfortable honest part
An agent this capable compresses schedules dramatically — eleven days for what would have been a quarter — but it does not compress judgment. Every hour I saved on implementation moved to decisions: what to charge, what to claim, what looks trustworthy, what to refuse. If you remove the human from that loop, you do not get a faster agency. You get a confident machine shipping stale exchange rates at scale.
That is exactly how we build agents for clients: production agents with evals, permissions, budgets, and human gates — because we run our own company on the same discipline.
The repository is public. Count the PRs.
FAQ
Common questions
Did an AI agent really build the NavyaAI website?
The agent wrote most of the code and copy across 34 merged pull requests between August 7 and 17, 2026 — the PR history is public on the repository. Every change was human-gated: plans approved before execution, designs approved from screenshots, pricing and claims decided by the founder. The agent executed; it did not decide.
Where was the AI agent wrong during the rebuild?
Concretely and instructively: it shipped currency display rates that were 5–6% stale until a verification pass caught them against live FX data, and it under-measured the site header by 5 pixels, leaving a visible white band that a human spotted from a screenshot. Both fixes shipped within hours because every change was verified against the live site, not assumed correct.
What kept the agent under control?
Gates. Plans were approved before code was written. Builds had to pass before merge. Every deploy was polled and screenshot-verified. Prices derive from one source file so displayed numbers cannot drift from the math. And the founder decided everything that was a judgment call: positioning, pricing, design taste, and every public claim.