miao.myRequests and approvals
Turn “request a monitor” into fields, an approval flow and a task.
- Read the request table and member roles
- Draft an approval plan
- Confirm before writing or notifying
MIAO AGENT · CAPABILITIES AND BOUNDARIES
Let the Agent handle work.
Let permissions set the boundary.
MIAO Agent is a controlled collaborator inside the server-side workspace: it reads resources allowed by the current app, proposes a plan, calls registered actions and waits for confirmation when needed. It is not a general-purpose agent that can run arbitrary code.
Demo data / illustrative Model names, records and receipts on this page are examples, not production capability or real customer data.
One inspectable run loop
01 DescribeState the goal and limits
02 ObserveRead authorized resources
03 DecideMake a plan and budget
04 ReviewShow impact, await review
05 ExecuteCall registered actions
06 ReceiptKeep events and result
Three practical work scenarios
Sales and service records
Create a follow-up summary or next task from customer records.
- See only fields visible to the user
- Call registered status-update actions
- Recheck permissions and record versions
Controlled data collection
Read credential-free HTTP(S) sources under declarative rules and prepare a draft.
- Declare source, fields and frequency
- Carry no browser token or private credential
- Authorize publishing or bulk writes
What it can read and do
Can read
Pages, fields, record summaries, versions, member permissions and historical run receipts within the current workspace and app authorization.
Can execute
Server-enumerated registered actions such as creating drafts, updating allowed fields, sending configured notifications or running declarative collection.
The harness/Jev selects legal actions; arbitrary tool names are not accepted.
Cannot execute
It cannot run shell commands, arbitrary user source, unauthorized HTTP or expand the creator’s business permissions.
Confirmation, cancellation, budget and durable runs
A plan shows impact, action count and budget first. Writes, releases, notifications or wider visibility enter “awaiting confirmation”; users can reject or cancel, while expired plans and changed permissions block execution. Authorized background tasks may continue after the browser closes and keep an auditable receipt; they do not retain browser login tokens.
One run receipt demo / illustrative
Illustrative run: customer follow-up · 2026-10-08
- Model (illustrative)
- demo-model-not-production
- Plan
- Read 24 → create 3 reminders → await confirmation
- Permission scope
- workspace:sales · app:follow-up · field:next_action
- Events
- Describe ✓ Observe ✓ Decide ✓ Review → pending
- Final receipt
- No write executed: user cancelled confirmation
- Budget
- demo limit: 20 actions · consumed: 3
Records and model names are illustrative.
What remains available when the model is offline?
Still available: deterministic pages in published apps, sign-in and permission checks, authorized CRUD, historical receipts, queued-run status and declarative collection results that do not depend on a model.
Waits: Agent plans to create or change apps, plus model-dependent summaries and suggestions. After recovery, permissions and record versions are checked again.
Not promised: this is not an open-ended general agent; an online model still gets no shell, arbitrary source, arbitrary HTTP or privilege escalation.