Claude turned Todos into Tasks. Let's test it out.

Kieran Klaassen and Trevan explore Claude Code's new task management system, comparing it to beads and discussing how it fits into the compound engineering workflow.

The Convergence

Everyone building AI coding workflows has converged on markdown task lists. Claude naturally updates them without instruction. Beads formalized this into a structured system. Now Anthropic has shipped native task tools that persist to .claude/tasks/.

The question: does this replace our custom approaches, or complement them?

Key Observations

Session-scoped by default. Tasks persist to ~/.claude/tasks/{session-id}/ and aren't automatically shared across sessions. You can set CLAUDE_CODE_TASK_LIST_ID=myproject to share, but it's a manual step.

Cross-agent potential. The real power is multi-session orchestration: one terminal as planner, another as worker, a third as tester. Each picks from the same task queue. This mirrors what people built manually with Beads.

Dependency tracking exists. Tasks support blocks and blockedBy fields, enabling proper sequencing. Claude can figure out what runs in parallel versus what waits.

Contextual Intelligence

Trevan raises the deeper insight: we've been forcing context into markdown artifacts when the better approach is letting machines query for context.

"I think this contextual intelligence is the next frontier... instead of forcing everything to create an artifact that the machine reads, how do we create a system in which the machine can query for context?"

He mentions Unblocked (getunblocked.com) as an example—a product that connects repos, Slack, and docs into a queryable knowledge layer. The agent asks questions; the system retrieves relevant context from wherever it lives.

Compound Engineering Implications

The compound engineering plugin already has todo tracking. The decision now: delete custom skills and rely on native tasks, or extend the native system with additional metadata (namespaced attributes like compound-engineering:priority).

Kieran notes a workflow gap: pending versus ready states. Sometimes AI proposes tasks that need human approval before work begins. The native system only has pending, in_progress, completed—no "approved" checkpoint.

Multi-Agent Pattern

The pattern they've found effective:

1. Planner session — Brainstorms and creates tasks
2. Worker session — Picks up tasks and implements
3. Tester session — Runs browser tests and validates

graph LR
    subgraph Shared["Shared Task List"]
        Tasks[(~/.claude/tasks/)]
    end

    Planner[Planner Session] -->|creates| Tasks
    Worker[Worker Session] -->|claims & completes| Tasks
    Tester[Tester Session] -->|validates & adds| Tasks

    Tasks -->|blocks/blockedBy| Deps[Dependency Graph]

With shared task lists, these can coordinate without manual handoffs. Each session maintains its own context window, avoiding the pollution problem of running everything in one session.

What's Missing

• No task deletion — You can't remove tasks, only mark them complete
• No subcategorization — Global IDs without project hierarchy
• No blocking status — Can't mark tasks as blocked by external factors
• Cross-project dependencies — The home directory approach doesn't scale to teams

Notable Quotes

"The value in the compounding loop isn't about task tracking and dependency management. It's really thinking, using judgment, taste, and the scaffolding above that."

"By having each task isolated into its own context window, you can now give it the ability to log any bugs for later."

Connections

• claude-codes-new-task-system-explained — Ray Amjad's technical breakdown of the same feature, with architecture diagrams
• beads — The inspiration for Claude's native task system; still offers features like hash-based IDs and stealth mode that the native version lacks
• ralph-wiggum-loop-from-first-principles — Another approach to fresh context per task, but without centralized orchestration