Claude Code Tasks: One Less Dependency

Steve Yegge built Beads to give coding agents memory. Tasks with dependencies, persistent state, multi-agent coordination. Then he built Gas Town to orchestrate 20-30 agents working in parallel. It works.

And now I’m watching Anthropic build the same architecture into Claude Code.

Beads solves what Yegge calls the “50 First Dates” problem: agents wake up every session with no memory. Markdown plans rot. Context conflicts. The agent can’t tell current decisions from obsolete brainstorms. The fix is a task graph—each task has dependencies, status, and an owner. Agents query what’s unblocked. State persists to git. Simple primitives, powerful results.

Look at the new TaskUpdate tool landing in Claude Code:

addBlocks: Task IDs that this task blocks
addBlockedBy: Task IDs that block this task
owner: Agent name for task assignment
status: pending → in_progress → completed

That’s Beads. And the recent changelog shows Gas Town patterns arriving too: launchSwarm to spawn multiple agents, teammateCount, team_name for scoping, mode for permission control.

Here’s where it gets interesting. Plan mode is becoming the entry point. You describe what you want. Claude builds a task graph—each task loaded with context, dependencies explicit. You review, approve, then launchSwarm spins up agents to execute in parallel, coordinated through shared task state.

Anthropic does this well: watch what works in the ecosystem, build it in. Beads proved the task graph pattern. Gas Town proved multi-agent coordination. Now the primitives you need are landing natively.

One less thing to install. One less thing to maintain.