How to track every AI agent your company runs

Claude, ChatGPT, and Codex can all run work on a schedule, and each ties that work to a person. Claude Code routines belong to one claude.ai account and act as that person. Claude's desktop scheduled tasks run only while the app is open and the computer is awake. In ChatGPT, scheduled tasks live in the account that created them, up to 15 active tasks on Pro and Enterprise plans, and sharing a task gives the other person a separate copy in their own account. Codex has automations of its own.

Add the cron jobs on laptops, the scheduled GitHub Actions workflows in individual repositories, and the Zapier or n8n flows with an AI step, and nobody has the whole list.

That's fine until something changes. The person leaves and the task stops, or keeps running with access nobody reviews. A tool changes its API and a report quietly goes wrong. The cost lands on a personal plan or an expense report. You can't fix what you can't see, so the first job is an inventory.

Keep the inventory short enough that people actually update it. A spreadsheet works for the first pass. What matters is that every automation has a named owner, that you know whose access it uses, and that someone notices when it fails or gets expensive.

The two fields teams most often skip are the accounts an automation acts as and what it can change. They decide what a mistake costs. A report that reads the CRM with a team account is low risk. A task that writes to customer records with one person's login is not.

A short survey finds most of them. Ask everyone the same three questions: what runs on a schedule or a trigger on your behalf, which accounts does it use, and what would break if it stopped tomorrow. Then check the usual hiding places yourself, because people forget the automations that work.

Don't switch anything off during the inventory. Some of the oldest automations are the ones a team quietly depends on, and the owner is the fastest way to learn what each one is for.

For anything that writes to a shared system, the team needs a record of each run that anyone can open: what started it, which tools it called, and what it changed. Personal tools keep that record in one person's account, and failures reach only that person, if anyone.

In type.com, each run of a managed agent automation makes a thread. Open it to see the agent's messages and tool calls. The automation's run history lists each run's status, time, deliveries, steps, and errors. For admins, Settings → Analytics → Health lists every automation in the workspace with its owner, Space, status, runs, AI cost, last run, and last error. It shows failing automations first. Select All to see the full inventory.

Know what the dashboard doesn't cover. type.com's docs say the Health view doesn't measure application uptime, missed schedules, queue delay, or integration sync freshness, and that no findings doesn't prove everything is healthy. For a deeper look, type-cli automations history lists failed runs grouped by reason, and can show the recorded tool and action steps of a single run.

AI costs hide in the same places as the automations. When people run tasks on personal plans, the spend is scattered across expense reports, and no one can say which automation is worth it. Put the work where the usage is metered in one place, then review it by team.

In type.com, the Cost tab shows type.com credits used, total tokens, cost per thread and per AI message, and the trend. You can break spending down by model, Space, channel, or person, and each Space opens a detail view with its costliest automations. A personal subscription coverage card shows how many members connected their own Claude or Codex subscription. The numbers are estimates; the bill itself lives in Billing.

For per-run numbers, admins can use type-cli automations costs, grouped by Space, automation, or run. It reports two views that you shouldn't add together: billable usage against the plan's allowance, and raw tokens valued at API rates, which also covers runs paid by personal subscriptions.

Limits apply to the workspace as a whole. Admins get a notice at about 80% of the monthly allowance, and at 100% new AI work pauses unless auto-upgrade is on. There's no budget per automation or per Space, so the monthly review is where you catch an automation that costs more than it's worth.

Start with the subscription. In type.com, each person can connect their own Claude or ChatGPT subscription. Agent turns they start and automations they create then run through it first, without drawing down the workspace's allowance, so moving a job to a shared Space doesn't have to change who pays for it.

Then recreate the automation. There's no importer for ChatGPT scheduled tasks or Claude routines, so copy the instructions into a thread in the right Space and describe the job. type.com proposes the automation with its trigger, schedule, destination, and connections, and a new automation stays off until you start it. Test it in a private thread before it touches shared data.

Bring the context along. type-cli sessions push copies a Claude Code or Codex session into a new thread in the Space, so the reasoning behind a script isn't lost, and type-cli skills sync uploads local skill folders so the shared version follows the same steps.

Scheduled runs in type.com use the creator's identity, model credentials, and billing context. That's the right default for personal work, and the reason team automations should use organization connections, such as the CRM account the whole sales team relies on, rather than one person's login. Admins can reassign an automation's owner when someone changes roles or leaves.

Decide what failure looks like before it happens. type.com pauses a scheduled instruction automation after three failed runs in a row, and those pauses show up under Needs attention in the Health tab. For the access side, see how to give AI agents access without exposing credentials, and for actions that need sign-off, human-in-the-loop approval workflows.

An inventory is only useful if it stays current. Fifteen minutes a week keeps failures from piling up, and a monthly pass keeps costs and duplicates in check.