On 8-17 you said 'move forward with them all' and the system ran 19 agents in parallel, one per priority, all finishing into staged drafts. Fan-out is the capability you reach for least and that pays a portfolio operator the most.
You juggle many ideas at once — that's the stated job. A single Claude working them one at a time is a bottleneck; subagents let one instruction become nineteen conclusions in an afternoon. You already own specialist agents and barely dispatch them on purpose.
A subagent is a fresh Claude with its OWN context window and tool set, handed a bounded task. It goes off, does the work, and reports back only its conclusion — not the pile of files it read to get there. That last part is the quiet superpower: your main thread stays clean because the searching, reading, and dead-ends happen in the subagent's context, and only the answer crosses back. For research-shaped questions ('where is X configured across the codebase?') this is why a subagent beats grepping in your own thread.
You have specialist agents wired already. standards-checker grades a finished artifact against the review rubric and returns JSON (for scripts to read); tom-review-judge grades against the SAME rubric and returns Markdown (for you to read in chat). Two agents, one rulebook, two output shapes — the pattern for any place you want both a machine gate and a human read. The 8-17 run was the extreme version: 19 general agents fanned out across 19 unrelated priorities, each isolated so one stalling couldn't block the others.
The discipline that makes fan-out safe is in the superpowers:dispatching-parallel-agents skill and one rule you already follow with tools: independent tasks go out in a SINGLE message so they run concurrently; dependent tasks must wait for their input. Fan-out is for the independent kind — nineteen briefs that don't need each other. A pipeline where B reads A's output is a different shape and must run in order.
You don't script this — you ask for it. The shape of the ask is the lesson.
# In a session, one message, three independent research questions:
# "Launch three Explore agents in parallel:
# (1) where is the reMarkable cooldown configured,
# (2) which LaunchAgents lack a witness check,
# (3) where does ink_apply write state."
# Each returns a conclusion; your main context never fills with file dumps.
# Your two judge agents share one rubric — see it:
sed -n '1,30p' "/Users/tom/Claude/JARBUS/_docs/review-standards.md"A subagent is a clean-context worker that hands back a conclusion — fan independent tasks out in one message, and reach for it any time an answer would otherwise flood your own thread.