If your Claude Code subagents suddenly stopped nesting, or a background agent kept spending past its budget cap, you weren’t imagining it. Anthropic shipped six real Claude Code changes across two releases in the past two weeks. Three of them hit the subagent system alone. Here’s each one, sourced straight from the changelog, and what to do about it.
- Subagent nesting flip-flopped twice in three days: disabled by default on July 21, restored at depth 3 on July 24.
- The budget cap had a real gap.
--max-budget-usdused to stop new subagents but let already-running ones keep spending. That’s fixed now. - Claude Opus 5 is the new default Opus model, with 1M context and fast-mode pricing — but Sonnet 5, not Opus 5, is still Claude Code’s overall default.
Claude Code Changes: The Subagent Limits Saga#
Three changes landed in the subagent system in a single week, and two of them contradict each other on the surface. Version 2.1.217, released July 21, added a default cap of 20 concurrently-running subagents. That’s a real safety limit. Before it, nothing stopped one message from fanning out an unbounded number of background agents.
The same release also disabled nested subagent spawning by default. A subagent could no longer spin up subagents of its own, unless you explicitly raised CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH.
Three days later, version 2.1.219 partially reversed that second change. Nested spawning came back, now allowed to a default depth of 3. If your workflows rely on subagents delegating further, your setup broke for three days and then quietly started working again. That’s exactly the kind of change that looks like a bug in your own code if you’re not watching the changelog.
If nested subagents matter to your workflow, don’t assume today’s depth-3 default is permanent. Two changes to the same setting in one week means it’s worth pinning CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH explicitly, rather than trusting the default to hold.
Every Change at a Glance#
Here’s the full set, matched to version, date, and the action worth taking. Each row is verified directly against the official changelog, not a summary of it.
| Change | Version / Date | Who it affects | Your action |
|---|---|---|---|
| Subagent concurrency capped at 20 by default | 2.1.217 — Jul 21 | Heavy parallel/fan-out subagent users | Set CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS if 20 is too low |
| Nested subagent spawning disabled by default | 2.1.217 — Jul 21 | Anyone with recursive subagent workflows | Check whether you relied on the old always-on nesting |
--max-budget-usd now halts already-running background subagents | 2.1.217 — Jul 21 | Anyone capping spend on background agents | Re-verify your cap — it now enforces mid-run |
| Nested spawning reinstated, default depth 3 | 2.1.219 — Jul 24 | Same group as above | Pin CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH for a specific depth |
| Claude Opus 5 released as the default Opus model, 1M context | 2.1.219 — Jul 24 | Anyone using Opus directly or via /model | Check if fast-mode pricing at $10/$50 per Mtok fits your usage |
claude-api skill defaults to Opus 5, migration path from Opus 4.8 | 2.1.219 — Jul 24 | Users of the bundled claude-api skill | No action, unless you’d explicitly pinned Opus 4.8 |
Bottom line: three of six changes are in the subagent system, and two of those three reversed course within the same week.
Budget Caps: What It Means for Your Bill#
The budget-cap fix is the one most likely to have cost someone real money before it shipped. Before version 2.1.217, --max-budget-usd had a gap. Once your budget was hit, Claude Code correctly refused to spawn new subagents. But any background subagents already running kept going, uncapped, until they finished on their own. Kick off a handful of long-running background agents right before hitting your limit, and they’d burn past it anyway.
The fix makes the cap do what it always implied. Hitting it now also halts already-running background subagents, not just new ones. If you run Claude Code with a budget flag set, this is worth knowing even though there’s no setup change required. The cap you already configured just got stricter, in a way that protects you.
If you haven’t set --max-budget-usd at all, and you regularly kick off background subagents, this is a good week to start. The enforcement gap that made it half-effective before is closed.
Opus 5 Is Now the Default — Should You Switch?#
Claude Opus 5 shipped July 24 as the new default Opus model, with a 1M-token context window and fast-mode pricing of $10/$50 per million tokens. It replaces Opus 4.8 as the default in both the /model picker and the bundled claude-api skill, which now migrates automatically unless you’d pinned the older model yourself.
Worth being precise here, because it’s easy to conflate two separate defaults. Opus 5 is the new default Opus model. Claude Sonnet 5 is still Claude Code’s overall default model — it shipped a few weeks earlier, on June 30, with its own 1M-token context and promotional pricing of $2/$10 per Mtok through August 31. Switching to Opus 5 is something you opt into on purpose. It doesn’t happen automatically just because you updated.
I’d reach for Opus 5 on large, single-pass refactors, where the 1M context window means the whole change fits in one turn instead of getting chunked. For everyday coding, Sonnet 5’s pricing keeps it the better default to leave in place.
How to Track Claude Code Releases Yourself#
None of this required guesswork. Every fact above came straight from the version-tagged changelog, and you can check it the same way. Run claude --version to see what you’re on. Then compare it against code.claude.com/docs/en/changelog, which lists every release in order, with the exact date and every change under it.
Claude Code ships new versions almost daily. Skimming the changelog every week or two, not just when something breaks, catches flip-flops like the subagent-nesting one before they surprise you mid-project.
The changelog, not a summary blog or a reaction video, is the only source that won’t be stale by the time you read it. Bookmark it if you run Claude Code in anything approaching a team workflow.
What’s Next#
Opus 5’s 1M context window deserves its own deeper look. Specifically: does it change how you’d structure a large refactor versus splitting it across turns with Sonnet 5, and what does the fast-mode pricing actually cost across a real multi-hour session? That’s worth a dedicated post once more real-world usage data exists.
Recap of the Claude Code changes that landed this cycle: subagent nesting flip-flopped twice in a week, the budget cap now actually stops running agents, and Opus 5 is the new default Opus model while Sonnet 5 keeps the overall default. Three of six changes touched the same subsystem — that pattern is the real story here, more than any single change on its own.
Which of these changes bit you first — the subagent nesting flip, or the budget cap gap?
Read next: Connect Claude Code to Any Model: 6 Exact Configs for the model-choice angle, and Run Claude Code for Free if the Opus 5 pricing has you rethinking your setup.
