Agent Flywheel guide: exhaustive markdown plans, beads tasks and agent swarms to cut token waste
doodlestein · x · 2026-09-16
doodlestein shares an end-to-end agentic coding methodology based on Jeffrey Emanuel's approach, via the Agent Flywheel stack.
- Core idea: move the hardest thinking into representations that fit in context windows — spend most of your time planning with a frontier model (e.g. GPT-5.6 Sol Pro in ChatGPT Pro), which produces a comprehensive markdown design doc.
- Task decomposition: beads-based ticketing turns complex projects into self-contained tickets, so each agent works like an assembly-line worker who only knows its own station — avoiding token burn.
- No subagents: every agent is a full harness instance; beads handles decomposition instead.
- On cheap models: they're 20-50x cheaper per token but burn 20x more tokens on hard problems; rigorous planning neutralizes that waste.
A full interactive guide with a flywheel simulation is included (metrics shown are illustrative, not benchmarks).
More from coding & agent
- Three agents citing one shared note is one lineage, not three verifications: on agent provenance design — RileyRalmuto · 2026-09-16
- User resolves flight issue in 15 minutes with Muse AI agent — armand_ruiz · 2026-09-16
- Gary Bernhardt: code review cuts agent diffs to 25%, "coding is solved" is hype — mgill25 · 2026-09-16
- Hooking a local LLM to GIMP via MCP: a working llama.cpp setup — BrianScottGregory · 2026-09-16
- Skillry Launches a Marketplace of Design-Focused Skills for Coding Agents — yihui_indie · 2026-09-16
- AI vet Moyix cringes at how OpenAI's Astra describes him to its subagents — moyix · 2026-09-16