Jaguarundi by XPayà
How this very website was built: a proof of concept in governing several AI models working as a team, coordinated by agents.
Context
This personal project is XPayà's own website. It started as a testbed for a very specific question: whether several AI models from different providers — not just one — can work on the same repository in a coordinated way, with explicit rules, without stepping on each other's work or each one deciding on its own.
The challenge
Two agents from different providers — Claude and Codex — working in parallel on the same working tree, with no human project lead checking every line. The real risk wasn't that the AI would write bad code, but that two agents would make contradictory decisions, overwrite each other's files, or that work would be lost if one of them ran out of runway mid-task.
My contribution
I designed the coordination protocol: a shared plan living in the repository itself as the only communication channel between agents, exclusive file ownership per work block, a mandatory validation phase before any implementation started, and a tie-breaking rule for when two agents disagreed. I kept for myself the decisions I don't delegate: approving the specification, deciding which names and data get published, and authorizing every commit and every deployment.
Decisions
I chose to keep the plan in a version-controlled file in the repository itself, not in any agent's memory, with dated, signed messages instead of silent decisions. No agent could make a unilateral call: every proposal needed explicit confirmation from the other, and ties were resolved by a third, disinterested model after two rounds without agreement. When one of the two agents ran out of runway mid-task, the other was able to pick it back up by reading the same shared plan, without losing context or repeating work — proof that it's the protocol, not any single agent's memory, that keeps the project running.
Outcome
The website itself is the result: trilingual content, accessible with no serious accessibility issues, meeting its performance budget, under a strict security policy, built entirely under this method. No wave of work was ever declared done with red checks, and every milestone was backed by a verified rollback rehearsal before moving on.
Scope and limits
This isn't a demonstration that AI builds a product on its own: every irreversible decision — approving the specification, which names get published, every commit, every deployment — stayed in my hands. What's being tested here is a multi-agent governance method, not a comparison of which model writes better code.
Technology
If this approach fits what you need, we can open a conversation.