Kidoka · 2026
A bug or request becomes a ticket. AI agents plan it, write it and review it. A person approves the pull request.
Before
Chat tools answer one question at a time. A person carries the ticket, the repository and the review in their head, and walks them from window to window.
The ticket carries the context. Agents hand it along, and the person steps in once: to approve.
The loop
Three roles
A planning agent, a writing agent and a reviewing agent. Each works from the ticket and a memory of the project, then hands the work to the next.
The pull request they open still waits for a person.
The core layer
Agents with roles, a queue of work and a memory of the project, inside a fence of tests.
React 19 · Vite · Tailwind · Express 5 · BullMQ · PostgreSQL · Docker · MCP
The fence
The tests are the fence the agents work inside: more than 3,000 of them, plus Playwright end-to-end runs.
0 / 3,000 passing · an illustration, one square per test
Speed
Kidoka went from a design doc to a working multi-agent platform in 12 days.
Now it builds itself.
Kidoka on its own repository
One arc per agent-opened pull request merged into main.
The one human step
Agents open the pull requests; a person merges them. 93 agent-opened pull requests merged so far.
Kidoka is built by ShengXing Chi, Head of Frontend at Know Your Customer, in Carlow, Ireland.