Anthropic’s Claude Code Adds Persistent Tasks to Turn Agents into Project Managers

Anthropic has updated Claude Code to introduce a persistent Task primitive that materializes multi-step engineering work as durable artifacts rather than ephemeral chat context. Tasks are represented as directed acyclic graphs written to the local filesystem, so agents can pause, resume, and be audited across sessions, machines, and runs. The primitive enforces explicit step dependencies so later actions cannot start until prerequisites complete, while environment variables let multiple agent instances point at and observe the same task set for parallel, role-separated workflows. Persisting state to disk makes plans survivable across crashes, includable in version control, and usable in headless, CI/CD, cron or pre-commit scenarios. Anthropic shipped stability patches alongside the feature—fixes for process-hang edge cases, CPU compatibility workarounds for SIMD-limited processors, and improved resume behavior across git worktrees—to reduce friction for enterprise deployments. Configuration knobs and opt-outs enable incremental adoption rather than a single global switch, letting teams migrate at their own pace. The update reduces pressure on model context windows by externalizing project roadmaps, but it also transfers new responsibilities to platform and security teams to secure artifacts, govern access, and expand testing and recovery policies. The broader ecosystem reinforces the need for those controls: complementary integrations from tools like Asana demonstrate how connectors bind models to real project data using OAuth and permissioning, provide two-way sync of work state, and often require explicit human approval for consequential changes. Those integration patterns highlight both the practical benefits—translating meeting notes and threads into executable project plans—and the operational trade-offs around auditability, admin visibility, and change-control. As organizations stitch Claude Code into existing Work Graphs and task systems, vendors and teams will need clear governance, role separation (e.g., generator vs. reviewer agents), and logging practices for unattended runs. The net effect is a technical step toward productionized, background subagents that can act, observe, and iterate with less human back-and-forth—paired with an organizational shift that demands new processes and controls to keep automation safe and auditable.