One queue for what needs you. Approve exactly what will happen.
Attention gathers what your apps have noticed into one list, ranks it and says why, lets a named approver approve the exact content, and keeps a receipt you can replay. It is a sandbox today: sample data and a simulated adapter, so nothing is sent.
Sample data What you can check on this page is a recorded run on sample data. A hosted copy of the sandbox is not running yet. The live demo page connects to a sandbox that you run yourself.
A recorded run
One item, from the signal to the receipt, taken from the sandbox. agentic-os v0.4.1, sample provider, simulated adapter.
The queue
12 items need a decision. The top of the list, highest first:
- milestone at risk: Production pilot high
- Visa documents missing — Taylor trip high
- thread (P0 lead) high
2 more are deferred, visibly, because they are beyond today's attention budget.
The item we follow
Visa documents missing — Taylor trip high
Departure in 8 days; the passport scan and visa form have not been received.
Evidence on the item:
- WhatsApp: "I'll send the docs soon"
- Booking TRIP-88: visa documents 0 of 2 received
- Signal seen
- Suggested from connected channels
- Approved by a named credential, bound to the exact content
- Sent through the governed path (simulated adapter)
- Delivery confirmed by an independent read
- Receipt recorded
Policy on this item: External message — approval is required. What approving does: Sends 1 whatsapp message to the recipients above. The receipt names the capability, personal_agent.connected_app_action, with status SUCCEEDED, written by the simulated adapter fake-external-agent.
Replay: what is checked
Replay reads the recorded trail and checks it. It does not run the action again.
- Evidence is attached to the item.
- The approval names the credential that gave it, not a name typed into a form.
- The approval is bound to the exact content that was sent. If the content changes after you looked, the approval is refused and you are asked again.
- The receipt and the policy decision refer to each other.
- Delivery was confirmed by a separate read, not only by the sender's own answer.
What is real and what is simulated
| What you asked for | What the sandbox does | What is simulated |
|---|---|---|
| Signals from several apps in one list | Merges signals from sample sources such as messages, invoices, quotes, support threads and project plans into one deduplicated list with stable ids. | The sources are sample records. |
| Evidence and provenance | Each item lists the records behind it with a reference, and the feed says what its basis is. | The records are sample records. |
| Prioritised recommendations | Ranks by tier and score and shows the reason and the policy version on every item. An attention budget defers the rest and says so. | The ranking policy is fixed in the sandbox. |
| Approval gates | Only an approver credential can decide. The approver is taken from the credential. A repeated approval does not send twice. | Credentials are throwaway demo credentials. |
| Execution receipts | Execution goes through the governed path and returns a receipt tied to the decision. | The adapter is an in-process fake. No message leaves the machine. |
| Replay | Reads the trail and checks it, as above. | It replays a sandbox trail, not a customer's. |
What it does not do yet
- There is no hosted copy. Anyone using it today runs the sandbox on their own machine.
- It does not connect to your accounts. Items that would need a mission to be approved are shown for review only.
- It has no customer results. We have not measured any; a pilot would report how often an approver could check the evidence and how long a decision took.
Open the live demo page How approvals work What is available today