Skip to main content

How many agents can you actually review?

Parallel agents multiply what gets written. What gets reviewed well is capped by your good sittings each day. Here are the two side by side.

The answer

You can keep 2 agents fully reviewed.

Lines a day: what 3 agents open, against what you can review well
Generated:
1,800 lines
You can review well:
1,200 lines

600 lines a day you can’t review well

The striped end of the top bar is the gap. Each day it either waits for tomorrow, so the queue grows, or gets a tired review that catches less.

Your numbers

Anthropic suggests three to five.

Assumption. Can be fractional.

lines

Assumption. Additions plus deletions.

Assumption. Each covers the lines per sitting below.

Assumptions
lines

The Cisco case study puts the limit at 400.

From one agent to ten

Each agent past the line adds the same amount to the daily gap.

Lines opened each day for one to ten agents, with your count in bold. The dashed line is the 1,200 lines you can review well; the striped part of each bar is what’s left over.
  1. 1 agent fits
  2. 2 agents fits
  3. 3 agents 600 over
  4. 4 agents 1,200 over
  5. 5 agents 1,800 over
  6. 6 agents 2,400 over
  7. 7 agents 3,000 over
  8. 8 agents 3,600 over
  9. 9 agents 4,200 over
  10. 10 agents 4,800 over

What to do about it

  • Run as many agents as you can review. Past that, the extra output either waits or gets a rubber stamp.
  • Make the history reviewable. The unit of history should match the unit of change. A pull request that needs three sittings is three changes, and reviewing it in one is a tired review.
  • An agent’s approval is not a human’s. A reviewing agent can catch a lot, but counting its approval as yours only hides the queue. It doesn’t clear it.

How this is calculated

What you can review well is good sittings a day × lines per sitting. What the agents open is agents × pull requests per agent × lines per pull request. The number of agents you can keep reviewed is the first divided by what one agent opens, rounded down.

Lines per sitting is sourced. A case study of code review at Cisco concluded that a review should cover under 200 lines, not to exceed 400, and take under 60 minutes, not to exceed 90. The default of 400 lines a sitting is the top of what that study calls good.

Three to five agents is a recommendation, not a measurement. Anthropic suggests three to five parallel sessions. Git isn’t what limits that number. Your review is.

Review doesn’t end at merge. A report on dotnet/runtime found that 52.3% of merged pull requests from the Copilot coding agent got commits from someone other than the author, against 10.3% of merged human pull requests. This page leaves that follow-up out, so the real number of agents you can keep up with is lower.

Pull requests per agent, lines per pull request, and good sittings a day are assumptions. Change them to match what you see. The page also leaves out more than one reviewer and anything live from your repository.

The other half of the trade: how much faster is fanning out?