Guide · 7 min read
AI agent time tracking: what to record and how
A timer is not the right instrument for an agent. An agent can run for four minutes and do an afternoon of research, or run all night and produce nothing. Still, the work happens on a client's task, someone asked for it, and sooner or later someone asks what it took. This guide explains what is worth recording about agent work, where to keep it and what not to do.
Why agents don't fit the usual timesheet
People's time is evidence of effort that nobody else can see. An agent's work is already visible: the request, the steps, the answer. Putting it on a person's timesheet inflates their hours and utilization for time nobody sat, and leaving it out hides work you may be entitled to bill and the signal about where AI helps.
The honest middle is to record agent time as the agent's own, on the same task as the people's, and never mix it into theirs.
Five things worth recording per agent task
- Who asked, and who stays responsible. The people on the task stay accountable.
- How long the agent worked, as finished duration on the task (not a ticking timer).
- What it cost: model, tokens and cost, once, on the task.
- What people did around it: the time to brief it and the time to review and rework its result.
- Who approved the outcome, a person or the agent itself.
A workflow that works
- Put the work on a task under the client's project, not in a loose chat.
- Hand it to the agent: assignee field, @mention or DM. The request becomes a session you can watch, answer and stop.
- The agent logs its finished time on the task, on the project's billable rules, when it's done.
- A person reviews it, moves the task to done, and tracks their own review time with a timer on the task.
- Reports show the agent on its own row next to people, billable if the task type is, and markable as invoiced.
Common mistakes
- Padding a person's hours with agent duration. It breaks utilization and trust.
- Logging agent time without a task. A spend line with no client task doesn't survive an invoice review.
- Counting only the model bill as cost and forgetting review time, which is often the larger part.
- Letting the agent mark its own work approved for anything client-facing.
- Billing for agent work the client never agreed to. Settle that in the contract first.
How Hourtick records it
Agents are workspace members with their own token or address. They take work over MCP, and log finished time with log_time on a task. Time follows the task type's billability and the project's rate and appears as the agent, for example "Claude (agent)". Agents never get a timer, and their time never lands on a person's timesheet, week total or approval. Reports also show how long each agent worked per request, and the tokens and model cost it reported.
Admins can scope each agent to clients and projects, make it read-only and cap its monthly cost or time, so recording is paired with limits.
Frequently asked questions
Should agent time be billed to clients?
If the client agreed to pay for the work, yes, shown transparently. Many teams bill it at the project rate, at a lower agent rate, or inside a fixed fee with the time kept as evidence. See the billing guide.
Why not just use a timer for the agent?
A running clock measures wall time, not effort or value. Logging finished duration on the task, plus cost and review time, tells you what the work took.
Does agent time count toward utilization?
Not a person's. Keep utilization about your people and report agent work beside it.
Which agents work with Hourtick?
Anything that speaks MCP: Claude Code, Codex, Cursor, Grok Build, Muse Code, cloud agents and your own scripts.
Is there employee monitoring?
No. Hourtick records no screens, apps or keystrokes. People track their own time.
Related guides
Track your first hour in a minute.
Free for your whole team, forever. $29/month when you need 5 GB.