Agent pipeline
From an idea to merged code, with AI agents and human review
A pipeline that runs personal software projects through four stages: flesh out an idea, specify requirements and design, plan build tickets with dependencies, and build them one at a time. A Python engine decides what runs next, keeps each project's plan, turns review comments into rework, tracks the monthly budget, and sends notifications over WhatsApp through n8n.
It drives a fork of the open-source agent-task-board, whose dispatcher runs Claude Code headless in a git worktree per task and opens a pull request; every change is reviewed and merged by hand. Routes fix each stage's model, cost ceiling and tools, and a sandboxed route lets agents build projects in any stack with no access to keys or tokens. In its first two weeks it produced 146 merged pull requests across six projects for $47.82 of reported model cost.
- Python
- Claude Code
- GitHub CLI
- n8n
- systemd
- bubblewrap
Screenshots
- How it fits together
- Four stages, each ending at a review
- The run record so far
- Guardrails for unattended agents
- One finished ticket on the board
The task board is David Chong's open-source agent-task-board (MIT), and the board screenshot is its own interface; the engine, routes, sandbox, prompts and workflows are mine. Costs are as Claude Code reports them per run.