Analytics

Use analytics to find patterns worth changing: repeated rework, rising usage, work that remains blocked, or decisions that still need frequent Human review. Open the underlying work before deciding what caused a pattern.

Look across the Workspace

Open Analytics in Pulse for work states, Human and Agent ownership, completions, approval evidence, Task cost, tokens, and active time. The Cost trend shows Task cost per day, or per week for ranges over 90 days, combining whole weeks for very long histories. The cost and active-time breakdowns split totals by stage, Agent, Project, Goal, Task, Task Type, and Worker, and the cost breakdown also lists every model.

Analytics has three views, and the filters and date range stay the same when you switch between them. Work shows progress, Tasks by state and ownership, completions, activity, and Autonomy. Cost shows Task spend, budget spend, the cost trend, the cost breakdown, and tokens. Time shows active time and parallel work.

Date ranges and daily and weekly charts follow your saved timezone. Today follows your local date, and weeks start on Monday. Changing your timezone refreshes the reporting without moving custom date values or clearing filters.

The date range applies to every card. On Work, Progress, Tasks by state, and Task ownership count the Tasks created in the range, and the Work item burndown charts the open work across the range, every seventh day for ranges over 90 days. Completed shows the Tasks completed in the range, one bar a day for up to a month and wider bars beyond, beside the change from the range before it. Activity is a calendar of the Tasks completed each day, shaded from light to dark by how busy the day was. It shows the last 52 weeks until you pick a range, then that range, up to 104 weeks, and lists the total, the active days and the longest streak of days in a row with a completion. It counts the Tasks that are currently completed, so a reopened Task leaves its day. Spend this period is the one exception, because a budget measures its own period rather than your date range.

Read progress and work-item burndown alongside the state and ownership breakdowns. More capacity may help a queue, but it will not fix an unclear brief or a missing handoff.

For a finite date range, the Parallel work multiplier compares summed active Agent execution with the complete elapsed period. Concurrent work adds together, so 1.4× means Agents recorded 1.4 times the selected period in task-attributed active execution. It measures parallel machine work, not equivalent Human labor. The peak follows the highest daily or weekly calendar period in the chart, while Project, Agent, Team, and Goal tabs attribute the same total without changing the denominator. All time omits this comparison because its elapsed period would be misleading.

If you can view budgets, Cost shows Spend this period against the combined limit of every active budget, split by stage. Each budget's own spend also appears on the Budget tab of what it covers.

Once Agents have asked each other something, Work also shows Agent requests. See the total raised, the share handled without a person, the share that reached a person, typical answer time, and the approval rate. Percentages have the exact counts beneath them, and each card compares with the range before it. You need permission to read Agents.

Every card has a history chart, grouped by the date requests were raised: daily for ranges up to 90 days, weekly for longer ranges. The charts show those requests' current outcomes, not a snapshot of what was known on that day. Typical answer time shows the median of the individual Agent answers in each bucket. Hover over a bar for its date and value.

Filter to what you care about

Use Filter next to the date range to narrow every section to Projects, Goals, Agents, Teams, or Workflows. In Agent requests, the Agent filter selects the agent that was asked. Filters combine like the Task board: choosing two Projects shows work in either, and adding an Agent then keeps only that Agent's work in those Projects. The filter stays in the page address, so you can bookmark or share a view.

Each filter follows the work it describes. Projects, Goals, and Teams follow the Task's own Project, Goal, and routing Team. Workflows follow the Tasks a Workflow created. Agents follow the Agent that did the work for cost and active time, and the Task's assignee for progress. With filters set, the budget spend covers only the budgets of the selected Projects, Goals, Agents, Teams, and Workflows.

Some cards appear only while their filter is set:

• Routing, under Work, shows how many Tasks the selected Teams routed, how often a routing decision was corrected, and how long human review took.
• Reliability, under Work, shows how often verifier checks of the selected Workflows passed, counting runs that match your other filters.
• Spend including Chats under Cost and Active time including Chats under Time add the selected Agents' Chats to their Task work. The rest of the page leaves Chats out because it measures Task work. Chats belong to no Project, Goal, Team, or Workflow, so they drop out once you filter by one of those.

Every card follows every filter you set. A breakdown hides the tab of a dimension you filtered to a single value, because it would show only that one row.

You only see filters for the kinds of records you can read.

Compare kinds of work

Choose Task Types in the cost and active-time breakdowns to compare Engineering, Marketing, Customer support, and other kinds of work. The token view also shows a Task Type breakdown.

Task Machine determines a Task's type from its brief during creation. A completed classification stays fixed even if the brief later changes. Unclassified includes Tasks without a completed evaluation and briefs that cannot be assigned confidently to a type. Task Types are automatic reporting context, separate from the labels you choose on individual Tasks.

The public Pulse page shows average recorded cost per completed Task by Task Type over the last 30 days. A type appears only when at least five distinct Workspaces contribute, and individual Tasks and Workspace names remain private.

Inspect your AI subscriptions

Choose Subscriptions next to Work, Cost, and Time to compare the usage of plans reported by your connected machines. The view and its Command Center destination, View AI subscriptions, appear after the first provider account is reported. The date range carries over from the other views, while Team, Workflow, and Agent filters do not apply here.

Task Machine comes first. Its month card shows usage against everything available this month on the same bar as the navigation usage meter, with one line for the forecast and when included usage refreshes. Its period card charts usage over the selected range. People who manage Billing can top up or change included usage from the card's menu.

Subscription logins stay on your machines, so Cloud runs never use them. Task Machine only reports provider credits such as Claude extra usage, Codex credits and an OpenRouter balance, and never enables, buys or spends them. Provider accounts follow, with the most-used plans first and more accounts loading as you scroll. Each account appears once, however many machines report it, and shows its reported plan, up to three of those machines, and its last successful reading. Accounts without a report in the selected range are hidden.

Each account has one card per usage window, plus any banked resets, extra usage, and balance the provider reports. A card's number is the latest reading in the range, and its chart shows the highest usage of every hour, day, or week, so limits you hit before a reset stay visible. When the range includes today, the chart continues with translucent bars for the usage expected until the next reset, and a weekly or monthly window whose current pace reaches its limit before the reset says when it is likely to run out, and the account is marked Runs out before reset. The forecast starts at the window's latest reset, including resets that came early, and names any banked resets that could cover the gap. Early resets in the range get their own card with the latest one's source.

A Codex subscription with ChatGPT credits gets a Credits card. Credits keep their own unit and never appear as money. When a limit is forecast to run out and the provider continues on extra usage or credits, the line says so and the Inbox alert says the provider bills the account owner instead of saying agents will pause.

When Task Machine Runs on a reporting machine used a subscription during the current period of a window that covers every model, its card shows how many tokens they used. Windows limited to one model show none. Once 5% of the window is used, it also shows how many tokens one percent bought. Cached reads are not counted. Use outside Task Machine, such as a provider's own app, fills the window too, so one percent buys at least that many tokens. Hover or focus the badge for the figures.

Once a window has two completed weekly or monthly reset periods, the account shows whether its plan fits, based on the highest usage each period reached: Upgrade plan when the limit was reached, or a banked reset was needed, in at least 30% of periods, Cancel plan when every window averaged under 25% and never passed 90%, Downgrade plan when every window averaged under 45% and never passed 90%, and Plan fits otherwise. The verdict follows usage only, not prices. Each card's subtitle says when the window is likely to run out, otherwise the usage expected at its reset, otherwise when it resets. When several windows reset at the same time, the account header also states that time. Periods without a reading stay empty rather than showing zero. A status badge appears only when an account needs attention, for example when a limit is reached, a sign-in expired, or a reading is out of date. Hover over or focus any badge to see how it was concluded: the rule that decided, each limit's numbers, and what to do next. These cards are read-only and do not spend credits or apply a reset for you.

While usage insights are on, people who manage Agents also get AI subscription alerts in the Inbox. Turning insights off closes open alerts.

• Running out: a weekly or monthly limit of a subscription an Agent uses is likely to be reached at least two hours before it resets. The alert closes when the limit resets or the forecast falls below 90%.
• Banked reset added and Banked reset expiring: a subscription gained a banked reset, or an unused one expires within 48 hours.
• Reset early: the provider reset a limit before its scheduled time.
• Plan fit: after at least four completed periods, the plan looks too small, too big, or unneeded. The same suggestion returns at most once every 30 days.
• Banked reset used automatically: a Codex subscription ran out of usage while an Agent or a tama code session was working, so Task Machine used the banked reset that expires next.

Each alert explains the numbers behind it, which Agents use the subscription, and what to do at the provider. Dismissed alerts do not come back. Choose which alerts also email you or notify your browser in notification settings.

People who manage Agents can rename an account to distinguish it from another plan. Renaming changes its Workspace label, not its provider login or plan. New reports update the view automatically. Workspace managers control collection and history in Privacy & data.

For a Codex subscription, people who manage Agents can turn on Use banked resets automatically from its menu or from a banked-reset alert. When that subscription runs out of usage and no other account in its pool can continue, Task Machine uses the banked reset that expires next so the Agent or your tama code session keeps working. An alert then tells you which reset was used and lets you turn the setting off. Only one session uses a reset at a time, none is used while Task Machine cannot be reached, and a reset is never used only because it is about to expire.

Inspect a Task

Open the Task to see its duration and usage cards. They help you distinguish execution effort, elapsed time, and the usage attributed to that work.

Use Run history for individual attempts and Browser use for recorded browser evidence. A failed attempt can still consume resources, so cost alone does not tell you whether the result was accepted.

Read periods and empty values carefully

Work, Cost, and Time follow the selected range. Spend this period follows each budget's own period. Read each card's label before comparing values.

An empty or zero value can mean that no usage was recorded for that scope and period. It is not proof that nobody worked on the responsibility.

An approval rate describes recorded decisions, not the percentage of every Task that was objectively correct. Keep learned autonomy separate from a simple success score.

Separate usage from funding and invoices

Analytics explain recorded work and usage. Operating budgets limit further work, while available managed usage determines whether Cloud work can continue. Payments, taxes, and billing adjustments belong in Billing.

See Duration and usage for effort measures and Budgets for operating limits.