Quick fixes AG-O002
Don't fan out past your review capacity
Shared methods · A shared method; linked tool guides explain the exact steps.
This page covers tools outside your selection. You can still read it. Find matching guides
Agents multiply what gets produced; nothing multiplies what you can actually read. The ceiling on useful parallelism is your review throughput.
My agents produce more than I can read? Here's the quick fix ↓ — the why sits right below it.
Quick fix: Don't fan out past your review capacity
Instead of Ten parallel agents, ten branches by lunch, and an afternoon of skimming merge buttons.
Try Run as many parallel agents as you can review today (for most people that is two to five) and queue the rest. Unread output is not throughput; it is inventory.
Why this works
Parallel agents shift the bottleneck rather than removing it: production scales with agent count, review scales with you. Past the ceiling, one of two things happens, and both are worse than going slower. Either work piles up unmerged and goes stale, or review quietly degrades to approval — and repository-mining research found exactly that in the wild: most AI-generated pull requests receive no human review at all, with stated review policies and observed behavior pointing in opposite directions. The fan-out felt like speed; the unreviewed merge is where its cost comes due.
When this applies
The generated-but-unreviewed pile is growing across days, or you notice your "review" of agent output has become scrolling to the approve button.
Template
Before spawning, one honest line:
Today I can properly review [N] pieces of this size. Spawn [N], hold the rest. Any output I will not read gets a critic agent before it gets my approval, not instead of it.
Go deeper
What real review of agent output looks like: reviewing agent code and reviewing Cowork output. The per-item version of this failure is output used unread; this is the same debt at fleet scale. Parallel mechanics without the collision risks: worktrees and parallel work.
Improve next
- Add a critic, not a committee
One adversarial critic beats a panel — smaller is stronger here.
- You are about to use something you have not read
Read it before you send it — the output is yours once it leaves.
Something wrong with this page?
Say what you expected and what you got. That is usually the shortest route to a correction, and it goes on the public issue tracker so the fix is visible.