Action Center

In one line. One screen that says whether everything is running, what needs a person, and where the time and money go — for one project, one workspace, or the whole platform. You’ll be able to. Open the Action Center at the right level, read each of its views, act on what needs you, set a monthly budget, and retry documents that failed in a run. Where this lives. Studio ▸ Governance ▸ Oversight ▸ Action Center for a project, Workspace settings ▸ Action Center for a workspace, and Account menu ▸ Admin ▸ Action Center for the whole platform.

Why it matters

A running solution throws off signals all day: documents that stall, runs that fail, spend that climbs, people waiting on a review. The Action Center gathers them in one place and ranks what needs a person, so nobody has to visit ten pages to learn that everything is fine — or that one thing is not. Every number on it opens the rows behind it and the page where you fix it.

The same screen at three levels

Level	Where	Who sees it	Its views
Project	`Studio ▸ Governance ▸ Oversight ▸ Action Center`	People whose role includes the Action Center page	Every view except Infrastructure and Data
Workspace	`Workspace settings ▸ Action Center`	Workspace admins	Every view except Data
Platform	`Account menu ▸ Admin ▸ Action Center` — Administration opens on it	Platform admins	Every view

The level decides what the screen counts, never a setting you can change on it. Anyone who can open it can read it. Acting — acknowledging an item, retrying documents, setting a budget — is for a platform admin anywhere, a workspace admin in their workspace, and, in a project, its admins and the people who may review or rescore there. Everyone else sees the numbers without the buttons.

The top of the screen

The status line — one sentence and a coloured dot: is everything fine, is something worth watching, or is something wrong.
The time control — Live (refreshes every 30 seconds while the tab is in view), 1 h, 24 h, 7 d, 30 d, Quarter or Custom. The choice follows you from view to view. With a window picked, each number is compared with the window before it; tick compare with the same weekday last week to compare like with like.
↻ reloads the screen. On the platform level there are also Checks — the list of everything the Action Center watches, where a noisy check can be muted or switched off — and Re-evaluate, which runs every check now.

The views

View	What it answers
Now	Is everything OK right now? Eight tiles — people, infrastructure, documents, projects, agents, data, cost, compliance — each one opening its own view, plus traffic and errors, a live stream of the last 24 hours, what changed, and a summary of what needs you.
People	Who is online right now, what they are doing and where; how many were on at once and the peak; who is new in this window; where people spend their time; and how fast pages load for them.
Infrastructure	Is the platform healthy? Its health and availability, the model providers, the next maintenance, how the services connect, CPU and memory, and the health checks. Not shown for a single project.
Documents	Is work flowing? The pipeline as a funnel and its throughput, stuck documents with Retry all, intake runs with their top failure cause and Retry intake, the ingestion jobs and connectors with their last pull and failures, who is reviewing right now, counts by stage, by collection, by day and by person, the time from received to processed, extraction quality, and the knowledge index.
Projects	Which projects are busy and which are in trouble: an activity grid, health by project, the biggest movers against the previous window, and what was created or installed.
Agents	What is running now, what is waiting on a person’s decision, runs and failures, who runs the agents, quality in production (the evaluation pass rate), the models and providers with their calls, errors, tokens and spend, one row per agent — and, in a project, the agents that did not run.
Data	For platform admins: database connections and slow queries, storage and cache, the heaviest database work, queues, connections to outside systems, and how long data is kept.
Cost	Where the money goes: spend over time, by model, anything unusual, Budgets, and what runs on its own in the background.
Compliance	Can we prove it? Availability of each capability, maintenance and incidents (Declare maintenance, Record incident), audit activity, guardrail blocks, budgets and data retention, access changes and API keys, and API service levels.
What changed	The activity record — sign-ins, edits, runs and access changes — with a count for each kind, a filter, and Export CSV.
Needs you	Everything that needs a person, ranked, with its count on the tab.
Ask	Ask the assistant about what the screen shows, in your own words.

Every number opens

Click any tile, bar or count and a panel opens beside it: What this means in one sentence, the rows behind the number (the first 50), What you can do here, and Where to act — a link to the page where the fix is made. A document row that came from a run can open that run step by step.

Needs you

Needs you lists every open item — a setup gap, a failed run, a budget about to be crossed, an incident — most urgent first. Filter it by Open or Parked, by severity and by source, or search it. On each item:

Investigate — opens the item’s next step: the view and the list behind it, or the page where it is fixed.
Acknowledge or Snooze 7 d — you have seen it, and it stops badging for a while. Snoozed is deferred, not done.
Assign — hand it to a person, by email, with a note.
Accept with a reason — a governance decision: the item stays on record, dimmed, with your name and why, and the check keeps watching.

Items clear themselves: when the signal behind one recovers, it leaves the list with no one closing it. A new check spends its first 7 days in a shadow week — it collects findings but does not raise them yet, and the screen says so rather than calling that an all-clear.

Recovering failed documents

A run that leaves documents failed raises one item per ingestion run, and the item stays open until they are recovered. It opens the list of those documents: each one’s run, the step that failed, the reason in plain words, and a category that says what to do — a temporary failure is likely to succeed on a retry; one that was over a limit, refused by the provider, or needs a setting fixed will fail the same way until someone acts. From the list:

Open run shows what happened, step by step.
Recover retries one document.
Retry the retryable first shows what a retry would do — how many would be retried and which are left for a person, and why — and nothing runs until you click Retry them now.

Only an admin of the project can recover its documents; anyone who can open the project can read the list.

Budgets

On Cost, Budgets sets a monthly budget in US dollars for the level you are on — and a workspace admin can also set one for each of their projects. The Action Center warns in Needs you before the month’s spend crosses it. Spend is an estimate, updated once a day. A budget item that is accepted with a reason stays accepted for the current calendar month only: on the 1st it asks again unless spend is back inside the budget.

Try it yourself

Open your project and go to Studio ▸ Governance ▸ Oversight ▸ Action Center.
On Now, read the status line, then pick 7 d in the time control and see which tiles moved.
Click the Documents tile. If anything is stuck, click the count and read What this means.
Open Needs you. Pick one item and click Investigate to see what is behind it; if it is expected, go back, click Accept with a reason and write why.
On Cost, set a monthly budget under Budgets and save it.

Recap

One screen at three levels: a project, a workspace, the platform. The level decides what it counts.
The top line says whether everything is fine; the time control sets the window for every view.
Twelve views: Now, People, Infrastructure, Documents, Projects, Agents, Data, Cost, Compliance, What changed, Needs you and Ask.
Every number opens the rows behind it and the page where you act.
Needs you ranks what needs a person; items clear themselves when their signal recovers.
Failed documents are grouped by run, and a retry is always shown before it runs.

Where to go next

G4 · Observability — run traces, and the platform console for performance and errors.
G5 · Audit & compliance — the full audit record behind What changed.
D4 · Ingestion & connectors — the jobs and connectors the Documents view reports on.
A12 · Crew & evaluations — the evaluations behind Quality in production.