From a YouTrack issue to an app deployed on staging. With no human involvement. An orchestrator plus 18 specialised AI agents, 18 pipeline steps, 12 autonomous workflows.
A YouTrack webhook starts the pipeline. It picks up a new Open issue and hands it over to the assigned AI agent — Claude Code, OpenAI Codex or Gemini.
Full REST API integration — reading issues and comments, changing states, publishing to the Knowledge Base. Newer comments take precedence over the original description.
The main Claude session drives the whole workflow. It first recognises the type of task, then delegates work to 18 specialised agents — it never implements anything directly, not even trivial changes. Tasks for Codex go through the same coder → reviewer → tester cycle.
Automatic clone, a branch named after the issue ID, commit, push and Pull Request creation. On an incremental re-run the existing PR is updated instead.
Every issue runs in its own Docker Compose project with an isolated environment. Automatic health check, with a fix and redeploy on failure.
Reverse proxy configured automatically over the API. A wildcard SSL certificate covers the staging domain. The proxy host is never deleted.
Loading the issue from YouTrack — description, requirements, comments. Choosing the workflow type and deciding whether the architect, designer, graphics, game graphics or voice-over agent is needed.
Issue → In Progress. The first visible action in YouTrack.
The task is classed as trivial, standard or complex. A small text change skips the design steps; a large change makes the architect mandatory.
Creating a dedicated working directory for the issue on the staging server.
Git clone of the repository and creation of a feature branch. The source branch from YouTrack is respected, otherwise the pipeline falls back to main/master.
If the project lacks up-to-date project documentation, the agent analyses the codebase and writes it (tech stack, structure, conventions, build). Every following agent loads it automatically.
Technical design — data model, API specification, dependencies, implementation order. Runs whenever non-trivial technical decisions are involved.
UI/UX design — layout, components, visual specification and design direction (palette, typefaces). It decides which graphic assets need to be generated.
Generates general graphic assets (logos, icons, illustrations) through OpenRouter. The cost is recorded in the run log.
Game graphic assets (sprites, textures, character art) through Leonardo AI. Used instead of, or alongside, general graphics when the project is a game.
Character lines are voiced through ElevenLabs and levelled to a consistent loudness. If a line fails, the pipeline carries on with subtitles.
Writes the code and tests and verifies the build. It receives the full context from the architect, designer, graphics and voice-over agents. For UI tasks it checks its own result on screenshots.
Code review — quality, security, conventions. Small issues in new code are fixed straight away, larger debt becomes separate tasks (step 15b).
Unit tests, build verification and a Docker image build — without starting any containers, so it never clashes with the real deployment.
A PR is opened on GitHub describing the changes. On an incremental re-run the existing PR is updated automatically.
The service starts in the issue's own isolated Docker Compose project.
Are all containers running, are the logs free of crashes and does the service answer over HTTP? On failure → coder → redeploy.
Nginx Proxy Manager publishes the service on the project's staging URL with a wildcard SSL certificate.
Screenshots of the running UI and a visual check against the assignment. The step is switched off for now — the coder checks the look itself and the E2E tester runs click-through checks on request.
What was implemented, the PR, the staging URL and how the agents performed. The orchestrator writes the summary itself for now; the Summarizer agent is temporarily off.
If the reviewer found technical debt and the workflow finished successfully, the orchestrator creates follow-up sub-tasks in YouTrack, tagged by the type of debt.
Issue → To Verify. A comment with the summary, the PR URL and the staging URL.
A record of the run in the central log — agents, results, iterations, token usage and cost. The record is validated before it is written.
Saving the state of the run (time, PR, staging, last commit, branch). This is what makes an incremental re-run possible.
Writes the project documentation — tech stack, structure, conventions. It passes that context to every following agent.
Claude Sonnet SUCCESS | PARTIAL | BLOCKEDTechnical design, data model, API specification. It analyses the codebase and hands the coder a plan and a map of the affected code.
Claude Fable SUCCESS | BLOCKEDUI/UX design, layout, components, visual specification and design direction. It decides which graphic assets are needed.
Claude Opus SUCCESS | BLOCKEDGenerates general graphic assets (logos, icons, banners, illustrations).
OpenRouter SUCCESS | PARTIAL | BLOCKEDGame graphic assets (sprites, textures, character art, pixel art).
Leonardo AI SUCCESS | PARTIAL | BLOCKEDVoices character lines, picks the best take and levels the loudness of the recordings.
ElevenLabs SUCCESS | PARTIAL | BLOCKEDThe heart of the pipeline. It writes the code and tests, verifies the build and checks UI work on screenshots. It consumes feedback from the Reviewer, the Tester and the Health Check.
Claude Opus SUCCESS | PARTIAL | BLOCKEDCode review focused on code quality, security and conventions. Small issues get fixed straight away, technical debt becomes separate tasks.
Claude Sonnet APPROVED | CHANGES_REQUESTEDRuns the unit tests, verifies the build and builds the Docker image — without starting any containers.
Claude Sonnet PASS | FAILClicks through the deployed app like a user — navigates, fills in forms and checks the resulting behaviour.
Claude Sonnet PASS | FAIL | BLOCKEDTakes screenshots of the deployed UI and checks the implementation visually against the assignment, contrast included.
Claude Haiku PASS | FAILThe final workflow summary. Until further notice the orchestrator writes it itself.
Claude Haiku SUCCESSWrites the application design — purpose, requirements, user stories, architecture, data model and UI concept.
Claude Opus SUCCESS | BLOCKEDInvestigates a reported problem, assesses its severity and decides whether it is a real bug. If it is, it drafts the fix as a task.
Claude Fable PROBLEM | NOT_A_PROBLEM | NEEDS_MORE_INFOProduces three visually distinct variants of a single screen as clickable HTML pages with previews.
Claude Opus SUCCESS | BLOCKEDBreaks a design document down into issues with acceptance criteria, dependencies and user stories.
Claude Opus SUCCESS | BLOCKEDOnce a day it reviews the whole workflow and proposes 1–2 concrete improvements as new tasks.
Claude Opus SUCCESSEvery day it finds new tips on working with AI and news from the AI world, each with a verified source.
Claude Sonnet SUCCESS | BLOCKEDThe orchestrator NEVER writes code, reviews, tests, designs architecture or UI directly — not even for a trivial one-line change. It always delegates to the proper agent. Implementing anything itself is a workflow violation.
Graphics & Game Graphics
Max 1 iteration
Coder ↔ Reviewer
Max 2 iterations
Coder ↔ Tester
Max 2 iterations
Coder ↔ Health Check
Max 1 iteration
Coder ↔ E2E Tester
Max 1 iteration
The full 18-step workflow from an Open issue to an app deployed on staging. Delegation to specialised agents, feedback loops, telemetry.
A shortened workflow for when the user sends an issue back for rework. It skips setup, clone, conditional agents and PR creation — the new comments are the main brief.
A larger feature is split into consecutive tasks that the pipeline works through one after another. Each task builds on the previous one; tests, the Pull Request and deployment happen once at the end.
Three visually distinct variants of a single screen as HTML pages. A person picks one in a comment and the chosen design is then implemented as a binding specification.
Investigating a reported problem without touching the code. If it turns out to be a real bug, a new task with a proposed fix is created.
Deploying an existing branch, changing a domain or diagnosing staging without any code change. As soon as a file in the repository has to change, the run switches to a full implementation.
A click-through test of the running app, acting as a real user. It runs only on request; on failure the coder fixes it and the test is repeated.
Automatic merge of verified issues into the auto-verify branch and chaining on to the next issue. Per-project locking lets different projects run in parallel.
Design documents written by the analyst, the architect and the designer, then published to the YouTrack Knowledge Base. Triggered by keywords such as "design", "specification" or "application plan".
Decomposition of a KB document into structured YouTrack issues with dependencies and a sprint plan. Larger capabilities are grouped into User Stories, which then run as a series.
A daily agent that reviews the current state of the workflow (agents, rules, tools) and proposes 1–2 concrete improvements as YouTrack issues.
A daily addition to the Knowledge Base: 2 tips on Claude/AI development and 1 piece of AI news. The agent may only read and search the web; the orchestrator writes the content after a duplicate check.
Every run is written as a single record into the central log. Token usage is filled in after the run by a deterministic post-processor; the monitoring dashboard reads from the log.
The state of the last run, stored with the issue — it enables the incremental re-run. When the pipeline finds it on the next start and it belongs to the same issue, the shortened workflow is chosen.
This happens when the review/test loop runs out of iterations without success, when the health check fails, on a critical blocker with no way forward, or when input from the user is missing. The issue receives a comment explaining the reason and the information needed.
After every agent finishes, a comment is posted to the YouTrack issue with a prefix matching the agent:
When Done issues are cleaned up, the pipeline automatically: