Anthropic Adds Persistent Context to Claude's Cowork Agent Threads

Anthropic rolled out persistent agent threads in Cowork on March 17, starting with Max plan subscribers and expanding to Pro within two days. The feature does something that sounds obvious but has been absent from most AI tools: Claude remembers what it was doing.

Cowork is Anthropic’s desktop agent platform, designed for multi-step tasks that span tool use, web browsing, file operations, and code execution. The persistent thread update means a task started on desktop Monday morning can be resumed from the iOS app Monday afternoon without losing context. Claude doesn’t need to be re-briefed. The conversation state, task progress, and agent context carry over.

This matters more than it sounds. The standard failure mode for agentic AI tools is context fragmentation — every new session starts cold, and the user ends up spending 20% of their time explaining the current state of a long-running task. Persistent threads eliminate that overhead. You launch a research workflow, come back four hours later, and Claude is where you left it.

The rollout sequence — Max first, then Pro — follows Anthropic’s pattern of using the higher tier as a beta population. Max is Anthropic’s highest consumer tier, with users who are more likely to run the kind of long-horizon tasks that stress-test persistence. By the time Pro gets the feature, the sharp edges are already filed down.

The mobile availability is strategically significant. Claude for iOS and Android now participates fully in the Cowork workflow, which means you can monitor an agent’s progress, redirect it, or hand off to a different device mid-task. That’s a meaningful step toward AI work that’s genuinely ambient — something running in the background while you’re not at a keyboard.

What Anthropic hasn’t addressed publicly is the infrastructure cost. Persistent context at scale — thousands of concurrent long-horizon agent sessions — is expensive to maintain. The selective rollout to paid tiers is likely both a feature test and a margin management decision. Free tier Claude users still get context resets between sessions.

For developers watching the agent platform space, the more interesting signal is how Cowork is positioning against Microsoft Copilot, Google’s Workspace AI, and Notion’s growing AI layer. Anthropic is building a cross-device, persistent, agentic interface on top of Claude. The question is whether developers can extend it or whether Cowork remains a closed first-party product.