Full autonomy where you allow it. None where you don’t. Decided per command, at the moment the command acts — on the exact application, process and window it is about to touch.
Everyone governs the data the AI reads. We govern the actions it takes.
Identity tells you who the agent is. A sandbox tells you where it runs. Approval tells you when it may. None of them tells you what it is about to touch. InvokeFlow checks that — every command, at the target, as it resolves right then.
Every command is checked against the exact application, process and window it is about to touch — at the moment it acts, never against something the caller supplied. If the target has changed, the command is refused.
A grant says where results may go. That destination is part of the command, and part of the record. A refusal is answered on your channel, never the caller’s.
Screen content, keystrokes, window titles and URLs have no field to land in. There is nothing to redact, because nothing was captured.
What was refused, and why, is kept in the same chain as what ran — including the moment someone changed the monitoring policy. Attested at the endpoint, where the action happened.
The broker is the agent’s only path to the desktop, and it runs in your environment. Revoke the grant — or stop the broker — and the agent can touch nothing.
A cloud key that opens one bucket, not the account. A sandboxed app that gets the file the user picked, not the filesystem. Data-loss controls that govern the channel. Authority bound to the action is how enterprises already stay in control — everywhere except the desktop, where the work actually happens.
Now it does.
Tamper-evident, hash-chained, and ready to hand to an auditor. Every command, every refusal and every policy change, from humans, automations, and AI agents alike, in one chain.
| # | Actor | Application | Command | Authority | Result may go to | Evidence | Outcome |
|---|---|---|---|---|---|---|---|
| 4181 | automation | CRM (desktop) | Record.Open | taught run · per-command | same application | read-back check | ran · verified |
| 4182 | automation | Dialer | Number.Load | taught run · per-command | same application | native response | ran |
| 4183 | agent | Sheet | Range.Read | grant-2f9a · per-command | same application | — | ran · text, 14 chars |
| 4184 | agent | Message.Draft | grant-2f9a · per-command | same application | — | ran | |
| 4185 | agent | Message.Send | grant-2f9a · destructive | external | — | refused · not named for this run | |
| 4186 | human | Message.Send | seat · human in seat | external | native response | ran | |
| 4187 | admin | — | Policy.Change | console | — | — | recorded |
No titles. No text. No URLs. A read’s result never enters the chain — a type-and-size descriptor stands in for it. Destructive steps are skipped unless a human names each one, every run: the agent’s send was refused, and the human’s ran. There is no “approve all.”
The runtime and the broker run in your environment. Nothing leaves the endpoint unless you send it.
The chain lives with you. It is evidence you own, not a report you request.
Overreaching. Taking access they weren’t given. Acting on a window that changed underneath them. The runtime treats each one as a refusal on the record, not an incident to investigate — and the way it checks authority is patent-pending.
It runs governed from the first command, with every refusal on the record — and the record is yours to hand to whoever asks.