AI agent time tracking records how long an AI coding agent, such as Claude Code, Codex or Cursor, spent working on a project, and credits that time to the person who launched the agent. It keeps the agent's time apart from the person's own hours, prices it at a rate of its own, and lets it reach a budget, a profitability figure or a client invoice like any other time, on lines that say what it was.
That one sentence carries three decisions, and each one exists because the obvious alternative goes wrong. This post explains them, what is actually measured, and how agent time differs from two things it is often confused with.
Why it exists
A developer at an agency starts an agent on a client's repository, describes a change, and moves on to something else while it works. Forty minutes later the change is ready for review.
If nobody records those forty minutes, the work disappears. The project looks cheaper than it was, a fixed-fee job seems to make more money than it did, and nobody can explain to the client how a smaller team delivered more.
If the developer records them as their own time, the opposite happens. Someone who runs two agents in parallel through an eight-hour day looks like they worked twenty-four hours, their utilization goes past 100%, and every capacity plan built on it is wrong.
Agent time tracking sits between the two: the work is recorded, and it is recorded as what it was.
The three rules
It belongs to the person who launched the agent
An agent did not choose the client, accept the brief or answer for the result. A person did. So agent time is recorded for that person, on the project they were working on, and they review it, correct it and answer for it. The glossary calls this supporting agent time, and it is the core of agent accountability.
It is not that person's own hours
Agent time is kept out of everything that describes a person: their timesheet total, utilization, capacity, reminders and the weekly summary. It still counts toward everything that describes the work: project budgets, profitability, uninvoiced time and invoices.
It has a rate of its own
An agent hour is not worth the same as a senior developer's hour, and it does not cost the same either. So agent time carries its own bill rate and its own cost rate, set for the workspace and adjusted per client or project where an agreement differs. Agent rates covers how to choose them.
What is actually measured
The useful measure is the time the agent spent working, not the time a session was open. A session left open overnight is not twelve hours of work.
So agent time is the sum of the agent's turns. A turn starts when someone submits a prompt and ends when the agent stops and hands back. The gaps between turns, while the person reads, thinks or is away, are not counted. Here is one afternoon:
| Time | |
|---|---|
| Session open, from first prompt to last | 3 hours |
| Four turns, the agent working | 50 minutes |
| Agent time recorded | 50 minutes |
There are two ways the time can be established:
- Measured. A plugin running beside the agent on the developer's machine times each turn. In Time Trakkr this covers Claude Code, the Codex CLI and Cursor.
- Checked in. An agent running in a vendor's cloud has no plugin beside it, so it reports when it starts and finishes, and the time is counted on the server's clock with strict limits that only ever count less. Checked-in time always waits for a person to accept it. See agent time in the cloud.
Neither records prompts, file contents or diffs. The record is the time, the project, the branch and a one-line summary of what the session did.
What it is not
Three different things get called "AI time tracking", and they answer different questions.
| What it records | The question it answers | |
|---|---|---|
| AI agent time tracking | How long an agent worked, for whom, on which project | What did this work take, and what should the client pay for it? |
| Token cost tracking | Tokens used and what they cost at list price | What did the model provider charge us? |
| AI-assisted timesheets | A person's own hours, drafted from their calendar or activity | What did this person work on today? |
Token cost is worth knowing, but it is a running cost, like hosting. It says nothing about how long the work took, and it is a poor basis for a client invoice. In Time Trakkr it is shown as an estimate beside the agent time, and never put on an invoice.
How agent time reaches an invoice
Agent time does not go straight onto a timesheet:
- A session is recorded only if its repository is mapped to a project and task. A session in an unmapped repository sends nothing, which keeps side projects and internal tools out of client billing. Admins map repositories with repository rules.
- It waits in a pending list that only its person can see.
- It is accepted, by its person, or automatically once the day has ended if the repository was mapped by an admin. The person can move it to another project first, or discard it.
- It is billed on its own invoice lines, labeled as agent time and placed after the person's own lines, at the agent rate.
Once accepted, an entry can be reduced but never raised above what was measured.
A worked example
A developer bills at $150 an hour. Their workspace sets agent time at $60 an hour to bill and $12 an hour as its cost. On Tuesday they work 6 hours on a client project, and the agents they ran record 2.5 hours on the same project.
| Hours | Billed | Counts toward their utilization | |
|---|---|---|---|
| Their own work | 6 | $900 | Yes |
| Their agent time | 2.5 | $150 | No |
| The project | 8.5 | $1,050 |
Their day is 6 hours, not 8.5. The project's budget is 8.5 hours lower, and its profitability counts the agent's 2.5 hours at $12 an hour, not at the developer's cost.
What Time Trakkr's agent time will not do
It measures time, not value: an agent hour that produced nothing useful costs the same as one that shipped a feature, which is why a person reviews it. It does not record anything from a repository nobody mapped. It never shows a manager what a session was about: managers see how much agent time each person has waiting, and nothing else. And nothing is recorded until an admin switches it on under Settings → Add-ons, though it is free on every plan.
Common questions
Is agent time tracking a form of employee monitoring?
No. It records what an agent did, not what a person did, and only in repositories an admin mapped to a project. It takes no screenshots and records no keystrokes, prompts or code. See why we don't build screenshots.
Should clients be billed for agent time?
That depends on the agreement. Some clients expect it on the invoice at its own rate, some agreements are fixed fee and it never appears, and some firms track it only to understand their costs. Agree it with the client before the first invoice that shows it. Billing clients for Claude Code and Codex compares four ways to do it.
Does it work for agents that run in the cloud?
Yes, by checking in rather than measuring, and that time always waits for its person to accept it. Claude Code cloud sessions can measure their turns from hooks committed to the repository.
How do I set it up?
An admin switches on Agent time under Settings → Add-ons (it is free on every plan), sets an agent rate and maps repositories to projects. Each developer installs a plugin with one command. Track Claude Code time per client walks through it step by step.

