AI Agents

An AI worker with one job, and only the permissions it needs.

Describe the job. SuNodes turns it into a clear job description, shows exactly what the agent could read and do, tests it, and only then lets it run. Anything that sends, publishes or changes something waits for your approval.

Creating an agent

Hire it like you'd brief a person

  1. Describe the job

    “Follow up on quotes we sent that haven't had a reply in a week.”

  2. Answer a few questions

    Only what this job needs: who it emails, which mailbox, how often. Values you type are used as you typed them.

  3. Review its permissions

    A plain-language review shows what it can read, what it can do, and what needs your approval. Nothing is created until you confirm.

  4. It's tested

    A test run reads for real but only checks its actions. It runs on its own only after the test passes.

What agents do

Agents, in practice

  • Hire an agent by describing the job

    Describe one job. SuNodes asks what it needs, shows exactly what the agent could read and do, and creates nothing until you confirm that review.

  • Schedules and events

    Agents run on a schedule, when a website form message arrives, when new email arrives in a mailbox it may use, or when you press Run now.

  • Approvals

    Anything that publishes, sends or changes something waits for your approval of that exact action; an agent can never approve itself.

    Limits: A mailbox owner can let one agent do some routine mail actions (replying in a thread, filing, marking read) without asking.

  • Tested before it runs

    Every agent runs a test first (reading for real, actions only checked), and runs on its own only after it passes.

  • Spending limits

    Each agent has a monthly credit limit; reaching it pauses the agent.

  • Connect your own services

    Connect a service that offers an MCP server or a documented web API, with its sign-in (including OAuth). You turn on each action an agent may use; reads run when allowed, changes need approval.

    Limits: There's no catalog of ready-made integrations: each connection is set up for your account.

Boundaries

What an agent may do on its own, and what it may not

Permissions are decided by policy, not by the conversation. The AI never sets its own permissions.

Can run without asking (when granted)

  • Read what it was given access to: your website, its analytics and form messages, public web pages
  • Prepare things for you: draft replies, recommendations, website edits to preview
  • Add records to your apps
  • Routine mail actions a mailbox owner has explicitly allowed (replying in an existing thread, filing, marking read)

Always needs your approval

  • Publishing website changes
  • Sending a new email (only the mailbox owner can approve it)
  • Deleting mail
  • Actions in your connected services that change something

Never allowed for any agent

  • Deleting a website
  • Changing domain ownership or settings
  • Buying services or spending money

Mail access is the mailbox owner's

  • An agent starts with no mail access
  • The owner grants each permission separately
  • Message content is anonymized before an AI model sees it
  • Card, ID and bank numbers, keys and verification codes are never revealed

Examples

Jobs an agent can do

Examples of what the agent design supports, not customer deployments.

Follow up on website leads

“When someone fills in our contact form, draft a reply for me.”

  1. Starts when a form message arrives
  2. Drafts the reply
  3. You approve before it's sent

Keep the inbox in order

“File newsletters and receipts every morning.”

  1. Reads anonymized mail it's been given
  2. Files messages, if the owner allowed it to
  3. Every action is recorded

Improve the website

“Suggest improvements to our site each week.”

  1. Reads the site and its analytics
  2. Prepares changes as previews
  3. Nothing is published until you approve

Questions

Agents, answered

Can an agent approve its own actions?

No. Approvals come from people with the right role, and each approval is for one exact action: if the action changes, it doesn't run.

What does an agent cost to run?

Agent runs use AI credits; what a run costs is shown in SuNodes. Each agent has a monthly credit limit that pauses it when it's reached.

Which services can agents use?

SuNodes's own (websites, apps, mail) and services you connect yourself through an MCP server or a documented web API. Each connection is set up for your account. See Integrations.

Can I approve from my phone?

The SuNodes phone app, which shows an agent's waiting action for you to approve or decline, is being tested with a small group and isn't in the App Store or Google Play. Today you approve in SuNodes on the web.

Hire your first agent.

Describe one job, review exactly what it may do, and approve what matters.