Kanban boards don't fit Loop Agents: agent observability should shift to outcome metrics
teortaxesTex · x · 2026-11-02
A thread argues that kanban boards are a poor fit for agents, especially loop-based agents: kanban worked because humans kept task topology and tacit knowledge in their heads, but agent workflows require making that knowledge explicit, overwhelming both the board and human tracking. Task lists in Cursor/Codex/Claude Code quickly drift out of sync with actual work.
The author proposes MCDT (Monotone Co-Design Theory) to structure collaboration between agents/sessions, with humans as input/output nodes who define goals and acceptance criteria while the loop plans its own dynamic todos. Key takeaway: agent observability should focus on outcome metrics around key fact nodes, not process metrics.
More from coding & agent
- "Someone must solve the cost of serving coworking agents" — hyhieu226 · 2026-11-02
- Open-source agent pays for x402 API calls with verifiable work instead of USDC — AlmostSneakers · 2026-11-02
- Dev builds game in 4 days with ~1,150 AI PRs, using TLA+ and lint rules to crush bugs — banteg · 2026-11-02
- Dev adds an agent to a smart switch — talk to it via Morse code taps — davidcrawshaw · 2026-11-02
- Claude autonomously runs 12-hour app migration, initial design would cost $3,000/month — tristanbob · 2026-11-02
- Behind the Craft lineup: OpenAI product VP, Claude designer, 10-person $32M AI company — petergyang · 2026-11-02