Claude Code release notes: 2.1.200-2.1.201 make agent questions manual

Claude Code 2.1.200 is not the flashiest kind of release. That is why it is worth reading. After several versions about background agents, draft PRs and agent queues, Anthropic changed how Claude Code waits for humans: AskUserQuestion dialogs no longer auto-continue by default. If the agent asks a multiple-choice question, the decision stays open until someone answers, unless you choose an idle timeout in /config.

The same release also makes permission language clearer. The old default mode is now labeled Manual in the CLI, --help, VS Code and JetBrains, and manual is accepted as a configuration alias. Version 2.1.201 followed with one narrow Sonnet 5 adjustment: sessions no longer use the mid-conversation system role for harness reminders.

Source: GitHub release v2.1.200, GitHub release v2.1.201 and the npm registry entry for @anthropic-ai/claude-code.

Claude Code release notes 2.1.200: questions wait for humans

A coding agent is an AI assistant that can read code, suggest changes and use tools inside a project. This update is not about giving the agent more tools. It is about when the agent should stop.

AskUserQuestion is Claude Code's tool for multiple-choice questions when the agent needs requirements, priority or a decision. In 2.1.198 and 2.1.199, the dialog auto-continued after 60 seconds of idle time. In 2.1.200, the default is that the question stays open. If you want an abandoned question to continue after a set time, you can set askUserQuestionTimeout to 60s, 5m or 10m in user settings or from the Question auto-continue timeout row in /config.

Source: Claude Code tools reference: AskUserQuestion.

That fits teams that actually use background agents. If an agent asks whether it should update the customer flow or only write a plan, that question should usually not resolve itself because someone went for coffee.

Manual becomes a clearer default mode

An approval gate is a point where a human must approve before an agent changes files, runs commands or affects a system. Claude Code already had a standard mode where tools ask the first time they need approval. Now that mode is described more clearly as Manual.

The docs say default still exists, but the CLI, VS Code and JetBrains show it as Manual, and manual is accepted as an alias from 2.1.200. For practical teams, that wording helps. Manual does not sound like failed automation. It sounds like an intentional working mode: read freely, propose freely, but ask for a decision when work crosses a boundary.

Source: Claude Code permissions: permission modes and Claude Code settings.

2.1.201: a narrow Sonnet 5 cleanup

2.1.201 is not a new workflow by itself. The release note says Claude Sonnet 5 sessions no longer use the mid-conversation system role for harness reminders. For most teams, the practical takeaway is simple: if you track Claude Code versions, 2.1.201 is current, but 2.1.200 is the larger user-facing signal in this pair.

Source: GitHub release v2.1.201.

What Nordic teams can do with the signal

This is a good week to write down how agent questions should be handled. Not a thick policy. A working routine people can follow when someone has three Claude jobs running at once.

Start with one focused workflow, such as code review, documentation updates or a CRM export turned into an internal action list. Decide which questions the agent should wait on, who owns the answer, when a timeout is reasonable, and what should appear in the run receipt. If the agent connects to real systems, use scoped API keys, environment variables or a secret manager, redacted logs, approval gates and audit logs. That is how integration becomes useful without making decisions disappear.

Try this prompt this week

Human step: Open the project or work queue where you already use Claude Code and check the version against the release notes. Do not update in the middle of a critical run.

Read our Claude Code settings, permission rules and latest agent jobs.
Compare them with 2.1.200-2.1.201: Manual as the default, AskUserQuestion without auto-continue and the latest Sonnet 5 cleanup.
Propose a question and approval routine for one workflow.
Show what the agent may do alone, when it should ask, and what belongs in the run receipt.
End with two changes we can make this week.

Good output should include:

• A concrete workflow, not a general AI policy.
• A clear rule for which agent questions must wait for a human.
• A timeout rule only where continued work is reversible.
• A run receipt with decisions, changed files or systems, and open questions.

Hammer angle

For Hammer, this is a Tool Forge signal: make AI agents useful in real workflows without turning every decision into a meeting. Manual as the default and questions that wait make agent work easier for colleagues to review, resume and trust.