Writing

How Many AI Agents Can One Person Run? A Formula for Your Number

By Scott Wilder · October 24, 2026 · 5 min read · All writing

You can run as many AI agents as you can brief and review in one sitting, and no more than your plan lets you open at once. For most solo founders with a day job that means two or three at first, four or five once your reviews stay clean. By the end of this post you will have a one-line formula that gives you your own number.

Your number is the smaller of two caps

Two limits decide how many agents you can run. Your number is whichever is lower.

  1. Review time. How many agents' output you can actually read, test and accept or reject in the time you have.
  2. Your plan. How many sessions your tool lets you run at the same time.

People who give a single number are usually describing their own setup. One practitioner says they can barely manage about 4 agents at once, and the point of that post is that the bottleneck moves from producing work to verifying and tracking it. A guide on the topic puts the common range at 3 to 5 parallel agent sessions per developer. Treat 3 to 5 as a reference point, not a target.

Size the count to your review slot, not to your ambition

Starting agents is cheap. Reading what they did is not. If you start a dozen at once, results and pending items flood in and you become the decision bottleneck. The fix is to design the count around a fixed amount of review time, which is the same idea behind sizing agents to the human's review capacity.

Here is the formula. Run it with your own numbers.

Review budget = minutes of review time per day ÷ minutes to review one agent's output

Example: say you have a 60 minute review slot each evening, and one agent's output takes you about 15 minutes to read and test. 60 ÷ 15 = 4 agents. If you only get 30 minutes after your kids are in bed, you run 2.

Round down. If the result is 3.6, you run 3.

Check your plan cap before you plan around it

The second cap is the number of sessions your plan lets you run at once. Reported Claude Code concurrent session caps are Free 1, Pro 5, Max 10. Plans change, so read your current plan page before you trust any number, including these.

Quota matters too. Parallel agents finish work sooner, but quota drains in proportion to how many run at the same time. Claude Code can also start several sub-agents on its own, and that uses tokens much faster than you expect. If you hit your usage limit early in the day, cut the parallel count or run some tasks one after another.

The final count, with two examples

Example: say your review budget is 4 and your plan cap is 5. You run 4. The plan cap is a ceiling, not a target.

Now say your plan cap is 2 and your review budget is 4. You run 2, and you queue the other two tasks for tomorrow.

Know the difference between a session and a subagent

These are not the same thing, and mixing them up costs you duplicated work.

The rule: help inside a task, use a subagent. A different task, use a new session. When you count agents for the formula above, count sessions. If your session spawns subagents, their output still lands in your review queue, so count that too.

Copy this task board to see your limit coming

When you lose track of who is doing what, you are already over your limit. Add a board before you add an agent. Keep it in a note, a spreadsheet or a text file.

Write your number at the top: Plan cap: __ / Review budget: __ / My number: __

Agent task Session or subagent Status (running, waiting for me, done) Next action for me
Fix signup email bug Session Waiting for me Test on staging, accept or reject
Draft help page Session Running None
Check broken links Subagent in the draft session Done Skim the list

Check it at the start and the end of your review slot. If more than half the rows say "waiting for me", pause new agents until that column clears. For tracking beyond this one board, see how to manage multiple AI agents without losing track.

Set it up in six steps

  1. Write each agent's job as one sentence with a stop point. "Done means the signup bug is fixed and a test passes."
  2. List the files, accounts and tools each agent may touch. If two lists overlap, merge the tasks or run them one after the other. Clear task boundaries are what let the count go up.
  3. Start with 2 or 3 agents on tasks that do not touch each other. Block one fixed review slot a day. If you are short on hours, fit the slot into a schedule you already have.
  4. Find your plan's session cap and write it on top of the board. Your number is the smaller of that cap and your review budget.
  5. Log per agent: task, accepted or rejected, minutes to review. Two weeks of this tells you your real review time better than any guess.
  6. After a week, add one agent or cut one. Add one only after a clean week of reviews. Cut one if reviews are piling up unread.

Treat the formula and the steps as a starting method to test against your own log. Your log wins over any number in this post.

Three mistakes that quietly cost you the most

Mistake What it costs
Starting many agents at once Results arrive faster than you can read them, and you become the bottleneck
Counting agents instead of review time Output piles up unchecked, so errors ship
Planning around the wrong plan cap Sessions get refused or queued, and quota runs out early

One more trap: going past 5 agents because other people say 10 or more. Some people claim higher numbers. If you want to go there, do it only for tasks that cannot break each other and that you can check in a few minutes each. Hard-to-check work on more agents just moves the pile to your review queue.

If you are not sure which tasks are safe to hand over in the first place, run the three question test for handing tasks to an AI agent before you add agent number four.

Your next ten minutes

Open your plan page and write down the session cap. Then time yourself reviewing the last thing an agent produced for you. Divide your daily review minutes by that time, round down, and take the smaller of the two numbers. That is your agent count for this week. Write it at the top of a blank task board.

Want a hand with the tasks you keep doing by hand?

If there is a manual task you think should already be automated, send me the details. I'm a solo builder, so you get me directly, with no handoff chain. Email Me with the task and what done looks like.

Questions

How many AI agents can one person run at once?

There is no fixed number. Your limit is the smaller of your daily review time divided by the minutes each review takes, and your plan's concurrent session cap. Most solo founders do well starting with two or three.

How many AI coding agents can I run in parallel?

A commonly quoted range is 3 to 5 parallel sessions per developer. Use it as a reference point, then size your own count with the review budget formula.

How do I know I have too many agents?

If more than half the rows on your task board say waiting for me, or reviews pile up unread, you are over your limit. Pause new agents until the queue clears.

Do parallel agents use up my usage limit faster?

Yes. Quota drains in proportion to how many sessions run at the same time. If you hit your limit early in the day, run fewer in parallel or run some tasks one after another.

What is the difference between subagents and parallel sessions?

Subagents run inside one session as scoped helpers with their own context. Parallel sessions are separate full instances on separate tasks, and they do not know what the others are doing.

Can a solo founder with a day job run a team of agents?

Yes, if the count fits a fixed review slot. Start with two or three independent tasks, log the minutes each review takes, and add one agent only after a clean week.