What happened. Anthropic releases Claude Code v2.1.212 with an important architectural change: /fork now copies the conversation into a new background session (a separate row in claude agents) while you keep working in the main session. The previous behavior—launching a subagent within the same session—has been moved to the new /subtask command. The release also adds limits to stop infinite loops (200 web searches per session, 200 subagents total, both configurable), automatically moves MCP tools running beyond 2 minutes to the background, and fixes bugs with worktrees, orphaned processes, and /ultrareview.

Why it matters to you. If you use Claude Code to delegate long tasks while working on something else, /fork now truly frees you: the forked session runs on its own, you can close the main window, and come back when it’s done. If instead you need a subagent that shares the context of your current session—for example for a quick check without losing your thread—you now call /subtask. The distinction is explicit: parallel delegation or in-line subtask. The anti-loop limits signal maturation: agents that run on their own need guardrails, and Anthropic is building them into the tool instead of relying on your continuous supervision.

If you want to try it. Update Claude Code to the latest version, open a project, and try /fork on a task you would have followed step by step before (for example: “read these five files, find inconsistencies, and write me a report”). Close the window and reopen claude agents: the forked session is still working. Then try /subtask for a quick check within the current conversation. The practical difference is when you review the result: immediately or later.

In detail

Context.

As we reported on July 15, Anthropic is shipping fast Claude Code releases: v2.1.208 brought accessibility and stability fixes, v2.1.209 fixed blocked dialogs in background sessions, v2.1.210 added live timers for long tools. Each release signals maturation toward daily use. This is the first one that changes the delegation architecture in a visible way.

What actually changes.

Before this release, /fork launched a subagent within the same session: it shared context, but you had to wait for it to finish before getting back to work (or open a new session manually). Useful for parallelizing, but not for delegating and forgetting. Now /fork copies the entire conversation into a new background session, which appears as a separate row in claude agents and runs independently. You can close the main session, and the fork keeps going. When you reopen claude agents, you see the status of all background sessions, and can resume one with /resume (which now opens a picker of past sessions, even ones removed from the list).

The previous behavior—subagent that shares the context of the current session—has been moved to /subtask. It’s the right command for “verify this thing quickly while I work on the rest”, where you want the subagent to see the same state as the main conversation.

The distinction is parallel delegation (fork) vs in-line subtask (subtask). Your choice depends on how long the delegated task is and whether you need to keep working without waiting.

Anti-loop guardrails.

The release introduces explicit limits: 200 web searches per session (configurable with CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION), 200 subagents per session (configurable with CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION). These are guardrails against infinite loops: an agent delegating to another agent that delegates to another, or searching in a loop for the same thing, now stops on its own. /clear resets the counters. It signals where the tool is heading: agents that run on their own need internal limits, not just external supervision.

MCP tools running beyond 2 minutes now automatically move to the background (configurable with CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS), so the session stays usable even if a tool is blocked or slow.

The fixes.

The release fixes a dangerous bug: plan mode was auto-executing Bash commands that modify files (touch, rm) without confirmation prompts and without calling the SDK’s canUseTool callback. Now it asks. It also fixes worktree creation that followed a committed symlink in .claude/worktrees, which could create files outside the repository. And it fixes SIGTERM during a Bash command that left the entire process tree orphaned in print/SDK mode: now aborting cleans up the tree and exits with code 143.

For /ultrareview, it now accepts PR references as #123, PR 123, and pasted URLs, and suggests the closest branch on typo. UX fixes, not functionality changes.

Practical implications.

If you use Claude Code for long tasks you used to follow step by step, /fork changes how you delegate: you can really let the agent work while you do something else, and check the result later. If you use subagents for quick checks within the current conversation, now you have an explicit command (/subtask) that says exactly what you want.

The anti-loop limits are necessary maturation: an agent running on its own without internal guardrails is a risk, and Anthropic is building those guardrails into the tool instead of relying on your continuous supervision. It signals the tool moving from “experiment to follow” to “tool to let work”.

The series of fast releases (208, 209, 210, 212) shows a pattern: Anthropic is smoothing UX for daily use, not adding features in bursts. Visible timers, dialogs that don’t block, tools that move to the background on their own, and now a clear distinction between parallel delegation and subtasks. Each release removes a friction point for people using agents every day.