Under a video walkthrough of a Grok Bot update, a commenter with 3 likes picks the feature worth keeping: "having one Page that shows what all your bots are doing, instead of checking each chat, makes a lot of sense." The comment ends with the better question: "Curious how it holds up once you have a lot of bots running at once."
The same video's comments add a second question: whether a dashboard like that uses extra tokens to run. And under a separate build video, viewers trade questions about the stack: what database holds the data layer, whether the page gets built with a coding agent. All three have plain answers. An agent's work leaves records in four places, a dashboard is one page that collects them, and the starter version is a document, not software.
Where the records live
An AI agent acts on its own: it runs steps in your other apps, on a schedule or mid-conversation, and it pauses when a decision needs a human. Work that happens while you're elsewhere leaves evidence in four places.
The conversation. An agent shows its work as it goes. ChatGPT's agent mode works on a virtual browser whose steps appear in the chat, and it pauses to ask before high-impact moves; the help page lists user confirmations among its safeguards. Screenshots the agent takes stay with the conversation until you delete the chat.
The task list. Recurring work has a home. ChatGPT's scheduled tasks are managed at chatgpt.com/schedules, where you can review, pause, or delete them, and OpenAI's troubleshooting for a task that stopped responding starts there: check for a paused task or a pending approval.
Notifications and approvals. ChatGPT's settings turn on push or email alerts for task runs. The approval prompt is worth its own check: an agent that wants to do something consequential usually has to ask first.
The results in your apps. The email sent, the file created, the calendar event added. The destination is where to verify the outcome, because the app shows what the work did.
Four questions per run
For each agent, four questions after a run:
What ran? The task list answers this: which schedules fired this week, which sat paused.
What changed? Look one level past the chat: the document edited, the rows added, the reply drafted.
What needs a decision? Pending approvals sit in the task list until you answer them.
What failed? The quiet failure is the one to hunt: a task that stopped without an error anyone read. The Scheduled page is where OpenAI says to look first.
The dashboard, two ways
A dashboard is one page that lists each agent's job, latest run, output, and pending decision. Two ways to get one:
Use the page the product ships. Some agent products build a status page in; the walkthrough the commenter watched showed one for Grok Bot. ChatGPT's version is the Scheduled page, plus chatgpt.com/schedules for recurring agent tasks. If you can't find yours, check Settings: notifications, app connections, and task lists live there.
Copy the starter version. One document, one line per agent:
Newsletter agent | sends Friday 9am | last run Oct 30 | draft written, waiting on my OK
Update the line when you check, or end the agent's instructions with a line like "write one sentence summarizing what you did to this document." The software version people describe in build-video comments (a database underneath, a custom page on top, a coding agent doing the wiring) can come later. The document adds no extra cost and doesn't depend on which product your agents run on.
What a dashboard won't catch
A dashboard shows whether work ran. It can't confirm the result was correct, and the failure it can't reveal is the quiet one: the task that ran and changed the wrong thing, or the automation that stopped between checks and sat unnoticed until someone needed it. OpenAI's own words on agent safeguards: they "don't eliminate all risks," and "it remains important to monitor ChatGPT agent."
The missing piece is a destination check. Pick one outcome per agent and verify it where the work lands: the newsletter draft is in the folder on Friday, the summary email went out, the numbers in the report match the source. Once a week, per agent, one check. That habit plus the page can make a second and third agent manageable, which is the commenter's scale question, answered.
FAQ
What is an AI agent dashboard?
A single page listing the status of the agents you run: each one's job, last run, output, and what's waiting on you. Some products ship one; a document with a line per agent is the starter version. The dashboard collects records that already exist (conversations, task lists, notifications, results); it doesn't replace them.
Does ChatGPT show what its agent did?
The main places: the conversation, where agent mode shows its steps and stores the screenshots it took until you delete the chat; the Scheduled page, which lists tasks and shows paused and pending-approval states; and the destination app, wherever the agent's output landed.
Do agent dashboards use extra tokens?
A page the product provides is ordinary app use; you don't feed it tokens. Token costs enter when a model reads or writes text: ending each run with a one-sentence summary adds a small amount of text to what the run already uses.
How do I know if an AI agent failed?
Check the task list first: paused tasks and pending approvals show there. Then check the destination: an agent that reports success while the app it works in shows no output is a failure the chat log won't fix. Notifications can catch many stoppages between checks.
Want to go further?
What Is a Personal AI Agent? Why the Demos Show Dinner Reservations covers the four consumer agents and where their approval controls live, and How to Automate Repetitive Tasks with ChatGPT is the walkthrough for the scheduled work this post tracks. If you want the cost side in more depth, What Is a Token in AI? The Unit Behind Every Price and Limit explains the unit. The one-step version of this post: open the Scheduled page for the tasks you already run and read what's waiting on you.