Buried in the July 21 release notes for Claude Code v2.1.217 is this one-line bug fix:
“Fixed--max-budget-usd not stopping background subagents: once the cap is reached, new spawns are denied and running background agents are halted."
Read it again. The flag whose entire job is to cap your spend was not stopping the agents you couldn’t see. Subagents started running in the background by default for everyone on July 1, the fix landed twenty days later. And the changelog never says when the bug began which is the unsettling part. For some stretch of those twenty days, possibly longer, the dollar ceiling you set governed the conversation in front of you while the fleet behind it kept working.
That single line is July 2026 in miniature. Anthropic shipped the engine first and installed the brakes after in public, one changelog entry at a time, 22 releases in 25 days. If you use Claude Code daily, those weeks explain every default that changed under you, and point at the settings that deserve five minutes of your attention today.
The engine arrived on July 1 in v2.1.198: subagents the helper sessions Claude spawns to delegate work now run in the background by default. Claude keeps working while they run and collects results when they finish, your conversation s only when a result blocks the next step. The fleet still asks before it acts background subagents surface every permission prompt in your main session but it no longer waits in front of you.
Two quieter v2.1.198 changes matter as much. The built-in Explore agent now inherits your session’s model instead of always running Haiku, so delegated research runs at your real quality tier. And subagents inherit your extended-thinking configuration before v2.1.198, they ran with thinking disabled no matter what your session did.
The pattern spread fast. By July 17 (v2.1.212), MCP tool calls running longer than two minutes moved to the background automatically, so slow servers stopped freezing sessions. By July 22 (v2.1.218), /code-review ran as a background subagent a review no longer fills your context window while you code and skills declaring context: fork backgrounded themselves by default.
Pro Tip: Anthropic paired the automation with restraint. Since v2.1.215, Claude no longer runs /verify or /code-review on its own, v2.1.218 added /deep-research to the manual-only list. Expensive skills now fire only when you ask which makes your token spend predictable again.
Delegation used to mean losing your terminal to the delegate. July fixed that with a new primitive: parallel sessions.
Since v2.1.212 (July 17), /fork [prompt] copies your entire conversation into an independent background session with its own row in claude agents, the CLI's dashboard of running sessions. You keep typing in the original, the fork tries the alternate approach. The old in-session behavior moved to /subtask <task>, whose result returns to your conversation when done.
Background agents also learned to finish the job. An agent that completes code work in its own isolated worktree now commits, pushes its branch, and opens a draft pull request instead of stalling on a permission prompt. The guardrails are specific: it never pushes main or master, never force-pushes or merges, and it skips the PR if you said not to open one. Fire-and-forget delegation finally returns something reviewable.
Before July: ask for a refactor, watch it happen, review, commit. After July: fork three approaches at 9 a.m., keep working in your main session, compare three draft PRs after lunch. The unit of delegation stopped being a task. It became a branch.
Now read the other half of the month, cap by cap. The quoted rationales are the changelog’s own words.
Read the sequence as a systems engineer and the shape is unmistakable: a capability ships, someone’s session fans out into something expensive or unkillable, and a bound appears within days. Nobody ships a 200-spawn cap unless somebody hit 201. This is load-bearing infrastructure being retrofitted with circuit breakers in production, with the receipts published.
If you run Claude Code unattended in CI, on a schedule, overnight that table is the difference between a bounded experiment and a surprise on your usage page. Three of July’s changes subvert reasonable assumptions.
Background subagents silently lose tools. A background subagent keeps every MCP tool but only a fixed list of built-ins Read, Grep, Glob, Bash, Edit, Write, WebFetch, WebSearch, and a handful more. Every other built-in is removed even when explicitly listed in the agent’s toolsfield, and the docs are precise about the failure mode: the removal reports no error unless it leaves the agent with no tools at all. The same definition resolves to different tools in the foreground and the background. A custom agent that depends on LSP has been running without code intelligence since backgrounding became the default. (AskUserQuestion never survives delegation at all it's stripped from every subagent, background or not.)
Warning: There is no per-agent way to force foreground. background: true exists in agent frontmatter, background: false does not skills got that opt-out in v2.1.218, agents didn't. Your only lever is the global CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1, which turns off allbackground execution reviews included.
Your permission globs changed meaning. Two changes in one week. v2.1.211 (July 15) moved “always allow” rules to the git repository root, so approvals persist across worktrees which is what you wanted. v2.1.214 (July 18) then fixed allow-rule matching: Edit(src/) now matches only <cwd>/src, where it previously auto-approved edits to a src/ directory at any depth, like packages/app/src/. That was a hole worth closing an allow rule shouldn't reach directories you never pointed it at but it means monorepo allowlists written before mid-July are now narrower than their authors intended. Rewrite them as Edit(/src/**) to restore the old reach deliberately. deny and ask rules kept their any-depth match.
Every subagent edit lands outside your checkpoints. This was always true, July made it matter. Edits a subagent applies aren’t captured by checkpointing, so /rewind can't restore them and subagents are now the default way work happens. The docs name the cases: a background /code-review --fix, any forked skill running in the background. The remedy is git, not /rewind. If checkpointing is your safety net, it now has a hole the width of the background fleet.
None of this is hidden, all of it is easy to miss. The cost of a platform rebuilding its execution model in three weeks is that your mental model from June is now wrong in places, and nothing forces you to notice.
July 2026 is when Claude Code’s unit of work stopped being the conversation and became the fleet background by default, parallel by design, bounded by caps that visibly arrived after the capabilities did. Remember the --max-budget-usd line not because the bug was severe, but because of what it marks: the month the platform's brakes had to catch up with its engine in public, over 22 releases in 25 days. The engine will keep outrunning the brakes. The changelog is where you watch the gap.
Do this today: open your .claude/settings.json and audit two things any single-segment glob rules like Edit(src/**) (they narrowed), and any custom agents that depend on tools outside the background keep-list (they're running without them). Ten minutes, two real bugs avoided.
What’s your concurrency cap set to and did you choose it, or did July choose it for you?
The Month Claude Code Became a Fleet and Got Brakes was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.