Agent time: the Agent work report and estimated cost
The Agent work report shows what people's coding agents did over a period, per person: accepted agent hours and the projects they went to, the most sessions a person had running at the same moment, the agent's own estimate of the work by hand beside the time the agent took, and an estimated token cost. Open it from the link in the pending agent time list on your timesheet, from the Approvals page, or at /reports/agent-work. It is not in the report tabs.
It covers accepted agent time only. Agent time that is still pending, or was discarded, is in no figure on it.
Who sees what
An admin or owner sees everybody.
A manager sees the people who report to them.
Anybody else sees their own totals.
The report is totals only. Nobody but the person who ran a session can see what it was: its repository, branch and summary are on their own timesheet and nowhere else. That is enforced in the database, not just on the page.
What each column means
Agent hours are accepted agent time entries in the period, with a line per project. They are not part of the person's own hours, utilization or capacity.
Most sessions at once is the largest number of that person's agent sessions that were running at the same moment, measured from each session's first turn to its last. A session that starts as another ends does not count as overlapping. It shows how much a person runs agents in parallel; it is not a number of hours.
Agent's estimate by hand is the agent's own guess at how long the work would have taken a person, added up over the sessions that gave one, beside how long the agent actually took on those same sessions. Sessions with no estimate are left out of both figures, and the report says how many gave one. Nothing checks the agent's guess. With no estimate from any session it says "No estimate given", which is not an estimate of nothing.
Token cost is explained below.
Estimated token cost
Token cost is what the tokens a session reported come to at a price per million tokens. It is shown on each pending session in your timesheet and per person in the report, and it is always an estimate:
It uses the price on the day you look. Nothing is stored on a session, so changing a price changes past estimates too.
It is the list price. A subscription, a discount, batch or priority pricing and long-context pricing are not reflected, so it is not what anybody was charged.
Cache writes are priced at the short-lived cache rate. A session that used longer-lived cache writes cost more than the estimate.
It is never on an invoice and has nothing to do with the hourly agent rate that agent time bills at.
Three answers are possible, and none of them is a silent zero:
An amount, marked "(estimated)". Different currencies are shown separately and never added together.
"no price set for" followed by the model's name, when a model has no price. That model is left out of the amount and never counted as free.
"not captured", when the agent sent no token counts. That means unknown, not zero. Cursor never sends them, and Claude Code and Codex send none when the agent's own files cannot be read.
Model prices
An admin or owner sets prices under Settings → Agent time → Model prices: input, output, cache read and cache write, per million tokens, with a currency.
The list starts as the vendors' public list prices for current Claude and OpenAI coding models, read from their price pages on 2 October 2026. They are not updated automatically.
Set your own price for a model to replace the default, or to add a model that has none. Remove goes back to the default.
A price is matched to the model name the agent reports. A dated or bracketed variant of a priced model (such as claude-haiku-4-5-20251001) uses that model's price; a different model whose name merely starts the same does not.
Leave a cache price empty to record that there is no price for it. A session that used that kind of token then shows "no price set".
Asking the assistant
The assistant can give you the same report for a period, with the same rule about who sees whom, and can list the model prices or set one for an admin. It cannot tell anybody what somebody else's sessions were.