Claude Code subagents hand off via files, not disallowedTools: Agent nesting

The article builds on a previous post that fixed the main Claude Code session in the role of PM, and now describes how that PM passes artifacts between specialized subagents. Three delegation options are laid out — natural-language prompting, @-mention, and launching the whole session with --agent <name> — but all three are ways for the human or main session to invoke a subagent, not for one subagent to @-mention another. So the author weighed two rejected designs against the adopted one: allowing nested delegation, which makes it hard to place a checkpoint before the PM proceeds, and merging everything into one giant agent, which loses specialization and bloats instructions.

The adopted flow runs market-researcher to product-strategist to builder to content-writer, with qa-reviewer recording a verdict in reviews/<slug>.md before publication. Artifacts land in research/, plans/, samples/, and articles/ or books/ according to each agent definition's description. The concrete payoff is the discrepancy between a source article's titled 43% and the body-counted 68%, caught by the PM and written into the plan as a risk item. Exactly the kind of review the author says direct agent-to-agent calling would have skipped.

The trade-offs are acknowledged: more PM instructions per step, no real-time parallelism for speed, and a need to predefine the directory layout, without which the author warns you regress to scattered artifacts. Whether this matters to a given team likely hinges on how much traceability and mid-pipeline review they need versus raw throughput, and the practical test is whether the extra PM overhead is worth the corrected figures. The author says a forthcoming book will cover generalizing the file-handoff design and when to use other integration patterns.