Ask an engineering manager what their team spent on coding agents last month, per person. You will get an invoice with one number on it and no names. Not because nobody looked, but because the per-person number does not exist anywhere they can reach.
Of the nine locally installed tools for running agents in parallel that we checked on 1 August 2026, none documents a spend cap. Of the twelve checked in total, two document one and neither is on by default. A category whose entire premise is running many agents at once has, so far, shipped no expense report.
The bill is a function of how many agents you run
This is arithmetic, not a discovery. One agent on one task costs some number of tokens. Five agents on five tasks cost about five times that, in the same wall-clock hour, which is the entire reason anyone bothers running them at once. Speed and spend are one lever with two labels.
The awkward part is the failure mode. The expensive event is not a bad diff, it is a loop. An agent retries a failing test, re-reads the same files, tries a variation on the same fix, and carries on doing that. One agent stuck like this costs you an evening. Five cost five, and at 02:00 nobody is watching either way.
What the twelve tools actually document
Two document a spend control, and the shape of both is worth knowing. Cursor asks you to set a spend limit the first time you start cloud agents, so the control exists but the default is a prompt rather than a number. Devin gives administrators per-user credit caps on top of a plan quota, which is the closest thing in the category to per-person control, and it applies to Devin only.
The one built-in budget we found in the locally installed group is amux's
AMUX_RATE_LIMIT_BUDGET, and reading its documentation carefully, it caps
automatic resumes after a rate limit, three per session per day. That is a
retry budget. It is not money.
Everything else is a vendor plan quota, which tells you the team hit a ceiling and never tells you who.
Why per-person is the number that matters
A total spend figure answers one question: can we afford this. Per-person spend answers the questions a lead actually has. Which work is expensive. Whether the engineer with the largest bill is doing the hardest work or has an agent stuck in a loop. Whether the number that doubled last month doubled because the team adopted agents or because one workflow went wrong.
There is a second-order effect that shows up on teams before anyone measures it. When spend is invisible, engineers self-police by guessing, and the guessing is uneven: the cautious ones under-use a tool the company is paying for, and the confident ones find out their run cost real money when somebody else reads the invoice. Neither is a policy. Both are the absence of one.
What Cohalen does about it today, and what it does not
Cohalen meters token spend per workspace. The limits layer is on in the default sandbox profile, the meter counts every turn, and a cap you set stops the next turn once that workspace crosses it. So a retry loop pauses and waits for you rather than running until morning.
Three limits, stated where the claim is, because a capability with hidden edges is worse than a smaller one:
- The cap stops the next turn. It does not abort a turn already in flight.
- The meter covers agents that report usage. Cursor publishes no per-turn token shape, so a Cursor workspace is not metered.
- An unset cap means the meter runs without a hard stop. The number is yours to choose, and choosing none is a choice.
For a team that wants to know what its agents cost before the invoice arrives, Cohalen is the tool building for that, and the check we ran says nobody else in this category has published an answer to it.
What you can do this week without any tool
Three habits get you most of the way while the category catches up.
Scope your prompts tightly. A task stated as one thing, with the command that has to pass, is the cheapest token control anyone has found, and vague prompts are what loops are made of.
Set a per-task ceiling in whatever you are using, even a crude one. A number you guessed beats no number, because the failure it prevents has no upper bound.
Name branches per task and per person. Then your git history is a rough spend ledger, and rough beats nothing on the day the invoice turns up.
Questions people ask about agent spend
How do teams track what their AI coding agents cost per person? Today, most cannot. Of the twelve tools checked on 1 August 2026, none reports spend per person. Two document a cap, and both leave the number to a user or an administrator to set.
What is the most expensive failure when running agents in parallel? A retry loop. An agent re-reads the same files and retries a failing test, and running five agents gives that failure five chances to happen unattended.
Does a spend cap stop a running agent? In Cohalen, the cap stops the next turn once a workspace crosses the threshold. A turn already in flight finishes.
Can you cap spend for every coding agent? No, and be suspicious of anyone who says yes. A cap can only meter agents that report their token usage, and not every agent does. Cursor, for one, publishes no per-turn token shape.