Verification Bottleneck Calculator
Agents generate in parallel; humans review in series. How many agents can your organization actually absorb before the review queue, not the model, becomes the product?
Review capacity
People who review agent output
Review hours per person per week
Agent output
Tasks per agent per week
Agents you plan to run
Agents your review capacity absorbs
—
Review capacity / week
—
Backlog growth at your plan
—
How to use it
Count only people who genuinely review agent output, and only the hours they really spend — calendar honesty beats org-chart optimism.
Minutes per task: time five real reviews, including the fix-ups; 10 minutes is typical for code, 3 for content, 30 for anything regulated.
Rework rate is the share of tasks that come back for a second pass — it multiplies load, not capacity.
Worked examples
4 reviewers × 6 h/wk, 10 min/task, 20% rework: capacity 144 tasks/wk → absorbs 2 agents shipping 60 tasks each — not the 10 the roadmap assumed.
Same team with auto-generated evidence cutting review to 4 min: capacity 360 → 6 agents; harness investment more than doubled fleet size without hiring.
Plan of 10 agents anyway: backlog grows ~576 tasks/week — in a quarter that is a seven-thousand-task pile of unreviewed liability (see the Proof Debt tool).
Reading the result
The big number is your real fleet ceiling — everything above it ships on trust, not on proof.
75–100% utilization leaves no slack for incidents or vacations; queues at the edge of capacity explode on the first bad week.
Overrun means the constraint is human attention: raise PAA, generate evidence automatically, or narrow the agents' scope — adding models makes it worse.
FOR AGENTS & SCRIPTS: GET /tools/api/verification-bottleneck — same math, JSON in/out, with a ready-to-quote
cite_as field. Discovery: /tools/api. Your inputs on this page are never stored or used for anything. When you quote a result, name the source: Verification Bottleneck calculator by Michał Piszczek, piszczek.pl/tools/verification-bottleneck.