Claude Code release notes: 2.1.211–2.1.212 bound agent workflows

Claude Code 2.1.211 and 2.1.212 make two practical changes to real agent work: they separate a new background session from an in-session subtask, and they cap searches and subagent spawns. Anthropic has also published its method for large code migrations. Together, these updates suggest a better working pattern: give agents substantial work, but keep the queue, the judge, and the stop points visible.

Claude Code release notes 2.1.211–2.1.212: clearer jobs, clearer bounds

In 2.1.212, /fork copies the current conversation into its own background session with a separate row in claude agents. The former in-session subagent is now /subtask. A background session is a full Claude Code conversation that keeps running without an attached terminal; a subagent handles one bounded part of the job inside the existing session.

The release also adds two default caps: 200 web searches and 200 subagent spawns per session. Long MCP calls automatically move to the background after two minutes. MCP, the Model Context Protocol, is the standard that connects Claude to external tools, databases, and APIs. Teams can adjust the caps with environment variables, but the useful default is straightforward: a loop cannot grow unnoticed simply because the agent can keep going.

Source: Claude Code changelog 2.1.212

Version 2.1.211 makes background work easier to trust. Claude should now report which agents are still running and wait for their real completion instead of filling in a result. A PreToolUse hook's ask decision also remains a required prompt in auto mode. An approval gate is a point where a person must approve the next action before the agent proceeds.

Source: Claude Code 2.1.211 on GitHub

Version 2.1.212 also fixes a concrete problem where plan mode could run file-modifying Bash commands without a permission prompt. That matters more than another convenient shortcut: a plan should remain reviewable before it becomes a change.

Source: Claude Code 2.1.212 on GitHub

Anthropic turns large code migrations into a controlled work queue

The second current signal is Anthropic's guide to code migrations with Claude Code. It describes ten migrated code packages and two larger examples: Bun's port from Zig to Rust and an internal port from Python to TypeScript. The useful part for most teams is not the line count. It is the process behind the work.

Anthropic starts with a mechanical judge that can test both the old and new systems. Next come a rulebook, a dependency map, and an inventory of gaps between the source and target environments. The team stress-tests those rules in a disposable run and throws the generated code away. Broad agent work begins only after the judge and the rules hold up.

Source: How Anthropic runs large-scale code migrations with Claude Code

The same pattern fits much smaller jobs: replacing a library, upgrading an API, moving a reporting pipeline, or rewriting a recurring integration. Do not start with more agents. Start with a testable referee and a work queue that can resume cleanly.

How to test this pattern in a real workflow

Choose a migration or refactor where you can measure the current behavior. Ask Claude to inspect first and edit later. When the job needs several tools, connect them through MCP with scoped permissions and credentials from a secret manager or environment variables. Put approval points before writes and merges, then retain test results and a run log.

This is a natural Tool Forge job: connect Claude to the right systems, then build the judge, permissions, and run receipt around the integration.

Try this prompt this week

Human step: Run claude --version and update to 2.1.212 through your normal installation channel. Open the correct codebase. Use /fork if the analysis should become a separate background session, or /subtask if it should stay inside the current session.

Source: Claude Code changelog and agent view documentation

Read the codebase and documentation for the planned migration.
Describe current behavior and the tests that can act as the judge.
Draft a rulebook, dependency map, and source-to-target gap inventory.
Propose one disposable stress test that challenges the rules.
Mark unknowns and the human approval point.
Do not edit any files yet.

Good output should:

• Separate verified facts from assumptions.
• Define a judge that can approve or stop the migration.
• Propose the next work item without starting to write code.

Once that holds up, you can fork the work, monitor the background sessions, and scale the agent count to match what the test chain can actually judge.