Claude Code’s New Task System Explained

Claude Code’s New Task System Explained

More

Descriptions:

Claude Code’s persistent task management system replaces the ephemeral in-session to-do list with a file-based task store saved in `.claude/tasks`, directly inspired by the open-source BEADS repository. Ray Amjad explains how this solves agent amnesia — the problem where starting a new session because the context window filled up meant the agent had to rediscover its place — by persisting task state to disk across sessions and subagents.

The system exposes four new tools available to both the main session and any subagent: TaskCreate, TaskGet, TaskUpdate, and TaskList. Tasks include dependency metadata, so some can be marked as blocked until others complete, enabling safe parallel execution. Amjad demonstrates this with a real workflow — adding ElevenLabs as a provider to his HyperWhisper speech-to-text app — and shows the measurable difference: using subagents with isolated task contexts consumed only 18% of the context window, versus 56% when running everything in a single session.

Two separate Claude Code sessions can now share the same task list, enabling genuine cross-session coordination without polling hacks. Subagents can also log bugs they discover as new tasks rather than dropping them due to a full context window. For developers working on multi-session projects, the task system mirrors real engineering workflows with handoffs and dependency tracking — and the orchestrator can instruct any subagent to annotate its task with information the next dependent task will need, creating a lightweight handoff protocol entirely within the Claude Code environment.


📺 Source: Ray Amjad · Published January 23, 2026
🏷️ Format: Deep Dive