Agent accountability for AI agent work you bill
Agent accountability means that every piece of work an AI agent does for a client has a named person who answers for it, and a record of what was done, before any of it can be billed.
An agency that bills a client for an hour of work can say who did it. When a coding agent did the hour, that answer gets vague: the agent has no name on the timesheet, its cost arrives as one bill from a model vendor, and the work tends to vanish into someone’s hours or never get billed at all. Agent accountability is the rule that closes that gap. It does not ask for less agent work. It asks that the work be owned.
The five requirements
A unit of agent work is accountable when it meets all five, in this order, on its way to an invoice.
- A person who answers for it. Agent work is credited to someone. Without that, nobody can be asked about a line on the invoice.
- A record of what happened. There is something to look at besides a total: when the work ran, where, and for how long.
- A chance to correct it. The person responsible can change the record before or after it counts.
- A rate somebody chose. Agent work is priced on purpose, not at whatever rate the person happens to carry.
- A label that reaches the invoice. Agent work stays distinguishable from human work all the way to the client.
How Time Trakkr meets them
In Time Trakkr, the work a coding agent does on someone’s behalf is agent time: supporting time that belongs to the person who launched the agent, not time of their own. It is an add-on an admin enables, for Claude Code and the Codex CLI.
- A person who answers for it. A coding agent’s session is recorded as agent time for the person who launched it, against the project and task its repository is mapped to. A repository nobody has mapped sends nothing at all.
- A record of what happened. Each session keeps its repository, branch, model, number of turns, start and end, and working time. Time between turns, and time spent waiting on a permission prompt in Claude Code, is not counted.
- A chance to correct it. Sessions wait in a pending list on the person’s own timesheet. They are accepted by that person, or automatically once the day ends for repositories an admin mapped, unless an admin turns that off. Either way its person can reduce or delete the entry afterwards, until it is invoiced and locked like any entry, and an entry accepted automatically is marked accepted automatically.
- A rate somebody chose. An admin sets agent rates for the workspace, a client or a project, each with a bill rate and a cost rate. With no agent rate, agent time cannot be accepted. It is never billed at the person’s own rate instead.
- A label that reaches the invoice. An accepted entry is marked Agent time on the timesheet, has a Source column in the Detailed report, and is billed on its own invoice lines labeled (agent time). It is kept out of the person’s own hours and utilization.
Next to the hours, the Agent work report shows an estimated token cost per person, so the cost of the work is visible beside its time. Managers see totals for the people who report to them; what a session was about stays on its person’s own timesheet.
Agents with nobody at the keyboard
Some agents are not launched by a person at a desk: a cloud agent, a scheduled job, a third-party agent connected over MCP. Their work cannot belong to whoever happened to start a session, so the accountability sits elsewhere.
- On your team, an agent member. An owner or admin adds it, with a name, a role, optional bill and cost rates, and the manager who approves its weeks. It tracks time under its own name, labeled Agent on timesheets, reports and invoices. Its entries need a start and an end, may not overlap, may not end in the future and may not pass 24 hours in a day. Disabling it stops its token and keeps its time on the books. How agent members work
- In a workspace an agent runs itself, an operator. Under the terms for agent-operated accounts, the operator is the person or company responsible for the agent, named at sign-up, who accepts the terms, and is the customer. The operator answers for the time the agent records, the invoices it issues and the data it puts in, including the work of any agent it adds. Time tracking for AI agents
What it will not do
- It does not read the work. Prompts, file contents and diffs are never sent. The record says how long and where, not what was written, and Time Trakkr does not judge whether the work was good.
- It does not guarantee a review. With automatic accepting on, which is the default, agent time from mapped repositories becomes time entries without anyone looking first. Turn it off under Settings → Agent time if every session must be reviewed before it can be billed.
- It does not invoice token cost. Estimated token cost sits next to the hours in the Agent work report, at list price on the day you look. It is never what was charged and never on an invoice.
- It does not cover every agent. Claude Code and the Codex CLI are supported. Cursor’s plugin has never been run against a real Cursor session, and agents that run in the cloud are not supported. There is no way to enter agent time by hand.
Questions about agent accountability
What is agent accountability?
Agent accountability means that every piece of work an AI agent does for a client has a named person who answers for it, and a record of what was done, before any of it can be billed.
Who is responsible for AI agent time in Time Trakkr?
The person who launched the agent. A Claude Code or Codex CLI session in a mapped repository is recorded as agent time for that person, and it waits in their own pending list. A manager or admin cannot accept or discard it for them.
Does a person review agent time before it is billed?
Only if your workspace asks for it. By default, agent time from repositories an admin mapped is accepted automatically once the day ends. An admin can turn that off so every session waits for its person to accept it. Either way, the person can reduce or delete an accepted entry until it is invoiced, and pending agent time never reaches an invoice.
What rate is AI agent time billed at?
An agent rate an admin sets, for the whole workspace or for a client or project, with separate bill and cost rates. Agent time is never billed at the rate of the person who launched the agent, and it cannot be accepted where no agent rate applies.
Can my client tell agent work apart from human work on the invoice?
Yes. Accepted, billable agent time is on its own invoice lines, labeled (agent time), after the person’s own lines. Agent time that is pending or marked not billable is never on an invoice.
Does Time Trakkr see my prompts or code?
No. For each session the plugin sends the repository address and branch, the model, turn counts and times, token counts, and the agent’s own estimate of the work when it gives one. It never sends prompts, file contents or diffs, and the service refuses a report that carries anything else.
Is the token cost what we were charged?
No. It is an estimate at list price on the day you look, shown next to the hours so the cost of the work is visible. Subscriptions and discounts are not reflected, and it never appears on an invoice.
Who answers for an agent that works with nobody at the keyboard?
On a team, an agent member is added by an owner or admin and can be given a manager who approves its weeks. A workspace an agent runs by itself has an operator, named at sign-up, who accepts the terms: under the terms for agent-operated accounts, the person or company responsible for the agent is the customer and answers for everything it does.
Read the details
The hours, the invoices, and what they earned
Free for unlimited people. One paid plan, $10 per member per month, when you want HR, scheduling, timesheet approvals, scheduled backups, email from your own domain and scheduled reports.
Start free