AI governance

AI that acts within authority you gave it.

In SuNodes, an AI being able to do something, being allowed to do it, and actually doing it are three separate things, each decided outside the AI. Here's how that works, and how it shows up when you build and run an agent.

Three separate questions

Capability, permission, execution

Can it?

Capability. What a tool technically does: read your website's analytics, draft a reply, publish a change. Each is described with its effect: read, propose or execute.

May it?

Permission. Whether this agent, for this job, is granted that capability. Grants are derived by policy from the job you approved, never written by the AI.

Does it, now?

Execution. Whether this specific action runs. Policy checks it again at that moment; anything sensitive waits for your approval of exactly that action.

Example: an agent can be capable of sending email, permitted only to prepare drafts, and so never execute a send at all.

How a decision is made

From request to record

How an action is governed An agent's request to act is checked against the permissions granted for its job; policy allows it, requires your approval, or denies it. An approved action runs once, exactly as approved, and is recorded. Requestan agent or theassistant wants to actPermissiongranted for this job?policy decidesApprovalyou approve thisexact actionActionruns once,as approvedRecordkept in theactivity log Not granted, or neverallowed: denied Declined or changed:doesn't run How an action is governed Requestan agent or theassistant wants to actPermissiongranted for this job?policy decidesApprovalyou approve thisexact actionActionruns once,as approvedRecordkept in theactivity log Not granted ornever allowed:denied Declined orchanged:doesn't run
Diagram of SuNodes's governed execution for AI Agents. Low-risk reads can run without an approval step.

The rules

What's enforced

These rules are built into how every AI Agent in SuNodes works.

Explicit authority

Every AI request runs at an authority level: observe, propose or execute. The effective level is the lowest of what was asked, what the job allows and what policy permits.

Permissions you review

Before an agent exists, you see in plain words what it can read, what it can do and what needs approval. The review is pinned: if the agent's job changes, it has to be reviewed again.

Approval of the exact action

An approval is tied to one action with its exact inputs, expires, and can be used once. An agent can never approve its own work.

Never allowed

No agent may delete a website, change domain ownership or settings, or buy services and spend money, whatever it's told.

Tested before it runs

An agent's first run is a test that reads for real but only checks its actions; it runs on its own only after passing.

Limits and a stop button

Each agent has a monthly credit limit that pauses it when reached, and any agent can be paused at any time.

Not everything asks first

What runs without an approval step

Approving every small step would make agents useless, so low-risk work runs once it's permitted. What counts as sensitive is fixed by policy, not chosen by the agent.

Runs when permitted

  • Reading what the agent was given (your site, its analytics, form messages, public web pages)
  • Preparing things for you: drafts, recommendations, website edits to preview
  • Adding records to your apps
  • Routine mail actions a mailbox owner explicitly allowed for that agent: replying in an existing thread, filing, marking read or as spam, unsubscribing

Always waits for approval

  • Publishing website changes
  • Sending a new email (approved by the mailbox's owner)
  • Deleting mail
  • Changes in services you connected

Traceable and revocable

You can see it, and you can take it back

Every run is visible

Each agent run shows its steps and what it did; decisions are recorded in your activity log. AI records keep fingerprints and short previews of inputs rather than full prompts.

Access can be withdrawn

Turn off a permission, pause an agent, remove its mailbox access or disconnect a service to take that access away. Mail access is checked again at every step an agent takes.

Your AI, or ours

Use SuNodes's managed AI, or your own provider key. Your key is never swapped for ours behind your back, and ours never falls back to another provider silently.

Where this comes from

SuNodes and Tavri

SuNodes is a product of Tavri, Inc., whose work focuses on governed AI execution. The governance described on this page is what's built into SuNodes; Tavri's wider research and technology aren't all part of SuNodes, and this page only describes what is.

Put AI to work, on your terms.

Hire an agent for one job, review exactly what it may do, and approve what matters.