Case study · Replay
Watch a real agent run
Three recorded runs of the local intake agent working the kanban project's own backlog — the same agent the AI-native workflow case study describes. Step through each: a note comes in, the agent retrieves related tickets, reasons about them, proposes an action, and a human approves or declines it.
These are real transcripts recorded against a local gpt-oss-20b — no live endpoint, no scripted output. The retrieval scores and token counts are whatever the run produced. The board in the middle is the same tickets the agent retrieved, and it rebuilds each run's change as you step.
A feature request nothing on the board covered.
The board, recreated step by step
Backlog
- Accessibility pass on the board UI
- Drag-and-drop tickets between columns
- list_tickets overflows the tool-output cap at scale
In progress
- Interactive replay viewer for agent runs
- Playwright e2e suite, path-filtered in CI
Done
- Local-first intake agent with retrieval
- MCP server exposes the board as typed tools
- Guard hook blocks dangerous git commands
- Executable definition-of-done gate
Incoming note
Add a keyboard shortcut that opens a quick-jump palette, so you can go straight to any ticket by typing part of its id or title.