Subagents made a Claude Code investigation 54% slower and 2.2x pricier, and Anthropic says fold them back
Subagent orchestration is not a free performance pattern; it multiplies token bills and wall-clock time on ordinary coding tasks and can turn code review into a 188x cost excursion for marginal gains. Teams should gate subagent use on main-context overflow or disposable parallel contexts, not on whether the work looks splittable.
Anthropic's Applied AI team now recommends collapsing specialized subagents back into the main agent loop. In a post by Diane Sarkis, one customer's switch to a single-agent loop ran about 2 times faster on the same tasks, and single agents with skills often beat both giant-prompt designs and subagent architectures on quality, cost, and per-task latency. Claude Code's own system prompt for stronger models was cut by more than 80%, partly because skills and tools changed to progressive disclosure.
A local benchmark in Claude Code puts numbers on that advice. A cross-directory investigation in a mid-sized Python/Swift repository took a main agent alone 93 seconds and $0.87 on average; splitting the same task across three parallel Explore subagents took 143 seconds and $1.92. Output tokens ran 3.6x higher and cache writes roughly doubled. All four runs found the same two key inconsistencies, so the split bought no quality.
Independent data on code review points the same way. A 20-PR comparison by Quickchat AI's Mateusz Jakubczak found the multi-agent /code-review max tier found two more severe issues than /code-review high, but cost $398.03 instead of $2.12, took 56 minutes per PR instead of 40 seconds, and produced twice as many extra reports per PR. Anthropic's own criteria keep subagents for three cases: measurable context-bloat slowdown, parallel tasks that each need a throwaway context, and tool-selection or instruction-conflict failures.
Subagent parallelism does not remove the serial dependency between reading code and merging conclusions; in the local run, output tokens grew 3.6x and cache writes roughly doubled, which is why wall-clock time went up instead of down.
Anthropic's 80% system-prompt reduction for Claude Code suggests that prompt design now delivers what orchestration was previously supposed to: keeping the main context small enough for the model to manage.
The Quickchat numbers are noisy and self-interested, but the margin is large enough that routine code review on maximum agent fan-out is hard to justify even before accounting for the vendor conflict.
The failure case Anthropic records, where a merged main agent answered from an order tool instead of asking the delivery specialist, shows that eval coverage must include production chains and inviolable rules, not isolated tasks.