I Built an Agent Team, Then Couldn't See What It Did
Description and use
Refmade is a desktop app that visualizes agent activity. Initially it relied on agents to write status reports, causing empty floors and uncertainty. Switching to log reading resolved the issue, enabling accurate real‑time team views.
Problem
The dashboard could not reliably show which agent was working because agents only reported status when they chose to, leading to empty floors and uncertainty about idle versus unreported work.
Outcome
Refmade now accurately visualizes agent activity by reading local logs, resolving the issue of empty floors caused by missing status reports.
Workflow
- 1. Define 11 subagent roles and a team lead in Codex.
- 2. Generate an initial office dashboard (v1) that displays pixel‑art floors based on agent status reports.
- 3. Observe that missing reports create empty floors and ambiguous idle states.
- 4. Decide to replace self‑reports with log reading.
- 5. Implement log parsing to extract session data and render team cards.
- 6. Update dashboard to show teams per session from logs.
- 7. Test with 25 Python and 16 Node auth/deploy tests.
- 8. Deploy Mac build 0.3.9.
- 9. Share public demo data and screenshots.
Evidence
The dashboard could not reliably show which agent was working because agents only reported status when they chose to, leading to empty floors and uncertainty about idle versus unreported work.
The first files of my agent team were written at 22:47 on September 20, 2026. By the next day I had a web dashboard to watch it.
The first version asked each agent to record its own state. A small recorder script wrote those reports to files and the dashboard drew them.
On September 25 I had decided the approach itself was wrong. How I replaced self-reports with log reading is the subject of post 3.
The current app draws a team per session from local logs instead of from reports.
- Record date: 2026-10-12T22:41:20.127Z