Four releases landed this week, and read together they tell one story: the long-running, walk-away job is becoming a first-class way to use Claude Code — and the app is finally giving you the safety rails and the cost dials to trust it.

Walking away got safer

The headline default change is quiet but it matters. Since 2.1.200, an unanswered AskUserQuestion no longer auto-continues. The old, undocumented behaviour let Claude wait about a minute on a question you weren’t there to answer, then pick a default and press on. Now it just waits.

For an overnight or unattended run, a stalled question is the safer failure mode — you come back to a paused session, not to a hundred commits built on a guess you never saw. Leave the new default alone for anything you kick off and walk away from. Only re-enable the idle timeout (via /config) for tasks scoped tightly enough that any reasonable default is harmless.

Then 2.1.203 removed the other reason you’d hover over a long run: switching back to the claude agents list used to silently stop a running subagent and re-run its prompt from scratch. That bug is gone — work carries over across the switch. The same release kills a 15–20s stall opening or switching background sessions on macOS. Between the two, background subagents went from “don’t touch it while it’s working” to something you can actually check in on.

Cost got a dial

2.1.202 adds a Dynamic workflow size setting to /config — unrestricted / small / medium / large — that nudges Claude toward a smaller agent count when it writes a workflow script. Set it to small before an ultracode request you’re unsure about, and watch the count in /workflows instead of learning it from the bill.

One caveat worth internalising: it’s guidance sent to the model, not an enforced ceiling. The runtime’s hard caps (16 concurrent, 1,000 per run) still apply, and a prompt like “audit every file in the monorepo” can still talk Claude past a small setting. Scope the prompt down too — the knob bounds the default, not the ambition.

The cheaper cost lever this week didn’t ship in a release at all. Simon Willison relayed a one-liner from the Claude Code team: tell Claude to push routine coding to a subagent on whatever lower-power model fits, and keep judgment, review, and synthesis in the lead model. He reports getting more done while his Fable allowance drained slower. Drop it in a session or your CLAUDE.md before a multi-file task — it’s a habit, not a config change, so it costs nothing to try and drop.

Also this week

  • Artifacts opened up. Publishing session work as a live page at a private claude.ai URL is no longer gated to Team and Enterprise — Pro and Max get it too, private to you. Stop pasting long PR walkthroughs into Slack; ask for an artifact and reopen the latest with Ctrl+].

One thing to watch

The centre of gravity is shifting from the interactive session to the supervised background run — and the tooling is following. Expect the next round of changes to keep hardening that path: better visibility into what a background fleet is doing, and firmer (not just advisory) ways to bound what it spends. The advisory workflow-size knob is a placeholder for a real ceiling; watch for one.