Fully autonomous pipeline
AI-powered pipeline from a ticket to a running app
Milan Knop · full-stack developer
I open a task in the tracker and AI takes it from there. An orchestrator splits the work between specialised agents — they prepare the design, write the code, review it, test it, deploy it and hand the result back into the task. Without me stepping in. I am a full-stack developer and this is the most interesting thing I am building right now.
// a single task
- Brief
- Design
- Code
- Review
- Deployed
Review → Code · max 2 rounds
How the pipeline works Send an e-mail
- 11 AI agents
- 18 pipeline steps
- 7 autonomous workflows
From an open task to an app running on a server.
Built with an AI-powered pipeline
// AI-Powered Development
A pipeline that carries a task from the brief to the server
Six phases, firm rules and a clear handover between agents. Design, code, review, tests and deployment all run without me stepping in — from an open task to an app running on a server.
-
Brief
A new task starts the pipeline. The agent reads the description and every comment — newer ones take precedence over the original brief.
-
Design
Depending on the type of task, an architect (data model, API), a designer (layout, components) or a graphic agent that generates assets with AI steps in.
-
Implementation
The coder receives the full context from the previous agents, writes the code and the tests, and verifies that the project builds.
-
Review
The reviewer watches quality, security and conventions; the tester runs the tests and the build. When something does not pass, the work goes back to the coder — for a strictly limited number of rounds.
-
Deployment
A pull request is opened, the app is built into a container, started and checked by a health check. A reverse proxy publishes it at its own address.
-
Handover
The task moves to the state awaiting verification and a summary waits in the tracker, together with links to the pull request and to the running app.
The orchestrator never writes the code itself. It always delegates to a specialised agent — even a one-line change.
- Review loop — max 2 rounds
- Test loop — max 2 rounds
- Health check — max 1 round
// What it runs on
- YouTrack
- Claude Code
- GitHub
- Docker
- Nginx Proxy Manager
- Playwright
- OpenRouter
- Leonardo AI
The whole pipeline in detail — all 18 steps, 11 agents and 7 workflows, including diagrams.
Open the presentation at ai-power.cz (in Czech) (opens in a new window)
// Roles
The agents that take turns on a task
Each agent has a single responsibility and returns a structured result to the orchestrator. Conditional roles step in only when the task needs them.
-
conditional
Architect
Technical design: data model, API, dependencies and the order of implementation.
-
conditional
Designer
UI/UX design: layout, components, visual specification, responsiveness and the assets needed.
-
conditional
Graphic agent
Generates images, icons and illustrations with AI models — including game assets.
-
required
Coder
The core of the pipeline. Code, tests, build. Takes feedback from every reviewing agent.
-
required
Reviewer
Code review: quality, security, adherence to conventions and detection of technical debt.
-
required
Tester
Runs the unit tests and verifies both the build and the container image.
-
more
…and more
An analyst for design documents, a sprint planner, a project documenter and an agent that improves the workflow itself.
// Who is behind it
From the data model to the last pixel
I am a full-stack developer — I build web applications from the data model and the API through a front end with forms and large data tables all the way to production. And alongside that I automate the development itself.
-
Front end
Components, forms and working with large volumes of data in tables. Accessibility, responsiveness, performance.
-
Back end
Data model design, REST APIs, integrations and running in containers.
-
Development with AI
Automating the development process: from a task in the tracker through review and tests to a running container.