The Beads Revolution: How I Built The TODO System That AI Agents Actually Want to Use

Posted twenty-four hours after the launch. Beads has gone from "an idea I was discussing with Claude on Wednesday morning" to a thousand GitHub stars and ~fifty forks six days later. The heart of the post is the diagnosis of why markdown plans don't work, in three bullets: (1) markdown plans are text, not structured data — parsing them steals model GPU cycles from the problem you're solving; (2) they're not queryable, so building a work queue or auditing forensics is hard; (3) agents rarely update them as they work, so plans bit-rot fast. The whole work graph has to be reconstructed from disk every time the agent asks "what should I work on next?" The rest of the post sketches what Beads gives you on top of that: epics with arbitrarily nested child issues, dependency edges, and a work queue agents can ask for the next ready item. The early-adopter community is already merging meaty PRs (race conditions, CGO removal, Windows support, an MCP server).

The cleanest one-page case for the design in the launch cluster: Introducing Beads has the origin story, Beads for Blobfish the civilian-grade explanation, Beads Blows Up the three-week verdict.