# Human Review Capacity Sets the Agent WIP Limit ## Claim Parallel agent work should stop expanding when completed runs begin waiting on human review. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - This applies when agents can execute independently but require a person for judgment, approval, or the next decision. - The exact cap depends on review complexity and context-switch cost; three waiting runs is one practitioner's threshold, not a universal constant. - Machine-checkable work can run farther without review, but only to the boundary established by its stop condition. ## Evidence [Human in the /loop](https://twitter.com/ericzakariasson/status/2070493377267646797/?rw_tt_thread=True) describes operating several loops in parallel and stopping new starts when three are waiting for review. The human reply becomes the next loop input. ## Links - [[The Bottleneck Moved From Code Generation to Review and Judgment]] — applies-in: cheap parallel execution makes human review the queue that controls throughput. - [[Heavy AI Use Removes the Fatigue Signal That Tells You to Stop]] — because a visible WIP threshold supplies a stopping cue that agent orchestration otherwise hides. - [[A Nondeterministic Check Breaks the Loops Stop Condition]] — extends: concurrency is safe only when each loop has a scoreable stop condition before it reaches the human queue. - [[3. Resources/Readwise/Articles/Getting Started With Loops — Claude Code Loop Design]] — applies-in: recurring code loops still need a review-bounded cadence. - [[3. Resources/Permanent Notes/Judgment at the End of an Agent Run Turns Every Late Failure Into a Restart]] — constrained by: the handover needs a point to interject before the run ends (Kemple). - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: the MOC lists it under judgment, review and taste as the new bottleneck.