# Human Review Capacity Not Compute Caps Parallel Agent Loops ## Claim The number of agent loops worth running in parallel is limited by how many decisions a human can review at once, not by how much compute or how many loops the infrastructure could technically sustain. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for loops that surface decisions or completions to a human reviewer. - Doesn't apply to fully autonomous pipelines with no human-decision gate at all — a different, riskier design this note doesn't endorse or evaluate. ## Evidence A practitioner running three to five long agent loops at once, plus shorter one-off agents: "If three are waiting on me I stop starting stuff and go review." Concurrency is throttled explicitly by his own review bandwidth, not by cost or infrastructure limits. ## Links - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — because applies-in; review capacity is the human half of the same maker/checker discipline. - [[3. Resources/MOCs/AI Agent Safety MOC]] — because applies-in ## Tests / falsifiers - A/B a capacity-throttled concurrency policy against a fixed-maximum policy and measure tasks completed per reviewer-hour; if fixed concurrency wins, the claim is wrong as a general rule. - Track whether review quality degrades measurably as queued decisions pile up past the reviewer's stated threshold. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/Human in the Loop (Eric Zakariasson)]] — states the concurrency-throttle rule directly.