Something changed in Claude Code over the last fortnight, and it was not the model. The interesting July releases were about control: how many helper agents you can run at once, how deeply they can nest, and how to run them in the background so they stop blocking your terminal. If you use Claude Code to actually build things, this is the difference between a tidy workshop and a room full of robots elbowing each other.
Subagents, in plain English
A subagent is a separate Claude instance your main session spawns for one focused job, with its own fresh context. You use them to run work in parallel and to keep the main conversation from filling up with noise. The trade-off is that a subagent starts blind: the only thing it knows is the instruction you pass it, so vague delegation gets you vague results.
The new guardrails
Anthropic spent late July fencing this in, and the changelog reads like someone watching agents get out of hand in real time. On 21 July, version 2.1.217 capped concurrently-running subagents at 20 and stopped subagents from spawning their own subagents at all. On 22 July, 2.1.218 moved /code-review to a background subagent and started running fork-context skills in the background by default. On 24 July, 2.1.219 brought nesting back at a sensible default depth of 3. The rule of thumb now: foreground subagents block the main chat until they finish, background ones run alongside you, and Claude Code asks for any tool permissions up front before a background agent starts, then auto-denies anything you did not pre-approve.
The /fork trick that saves real money
Here is the one worth learning. A fork is a special subagent that inherits your entire conversation history instead of starting cold, so it is perfect for trying several approaches from the same point. You call it inline, for example /fork draft unit tests for the parser changes so far. It is still experimental: you need Claude Code v2.1.117 or later and the environment variable CLAUDE_CODE_FORK_SUBAGENT=1. The payoff is not just convenience. Forked children share their parent’s prompt cache prefix, which cuts input token costs by up to 90% for the second child onward. If you are running three variations of the same task, two of them are close to free on input.
What to actually do
Do not confuse a higher limit with a good idea. Twenty parallel agents is a ceiling, not a target, and every one you spawn is context you have to reason about. Use background subagents for reviews and test runs while you keep coding. Reach for a fork when you want to compare two or three real attempts rather than argue with one agent about which is better. And whatever you delegate, hand over only the context that job needs, because a subagent cannot see your screen, only your instructions. Drive a small, well-briefed fleet and it feels like leverage. Spawn twenty vague ones and you have just paid to generate your own cleanup job.
Did you know: the term “debugging” predates computers by decades, but the 1947 log entry of an actual moth taped inside a Harvard relay is why engineers still say they are pulling bugs out of the machine.