Skip to Content
SafetySafety

Safety

You’re always in control of what your assistant can do. Town provides multiple layers of safety controls so your workflows behave exactly as you expect.

Safety controls

The basics

Permission options

When you’re chatting with your assistant, you can control how much autonomy it has for that session:

SettingWhat it does
Ask before any changesDefault. Every action that makes a change requires your approval first.
Always allow safe actionsAutomatically approve low-risk actions like labeling and archiving. Actions that communicate externally still require approval.
Allow all actionsFull autonomous mode — no approvals required for this session.

These settings reset when you start a new chat.

Workflow modes

Every workflow has a default mode that applies whenever it runs:

ModeWhat happens
AutonomousActions execute immediately without asking
Approval-requiredActions that make changes wait for your approval
Read-onlyCan only read and analyze — no modifications allowed

Per-tool controls

You can set different modes for individual tools within a workflow. For example, a workflow might be autonomous for labeling and archiving, but require your approval before sending any emails.

Action logging

Every action your assistant takes is logged. You can review what happened, when, and why — and adjust settings if needed.

Progressive trust

A common pattern is to start restrictive and loosen over time:

  1. Start with Read-Only — See what the workflow would do without any risk
  2. Move to Approval-Required — Let it propose actions for your review
  3. Promote to Autonomous — When you’re confident the behavior is correct

Start with approval-required for new workflows. Move to autonomous after you’ve reviewed several runs and confirmed the behavior is correct.

Built-in protections

Town includes security measures that apply automatically:

  • Smart safeguards — When a workflow has access to private data, untrusted input, and external communication tools, extra protections kick in automatically
  • Input validation — Blocks known manipulation patterns in incoming content
  • Output validation — Checks tool arguments before executing actions
  • Account isolation — Workflows can only access accounts you’ve connected

Learn more in Security.

Best practices

Start narrow, expand carefully. Begin with minimal tools and permissions. Add more as you verify behavior.

Review run history regularly. Even autonomous workflows should be reviewed periodically to make sure they’re doing what you expect.

Use approval-required for outgoing messages. Sent emails and messages can’t be unsent. Always review before sending.

Test before going live. Use read-only mode to verify workflow behavior without making real changes.

Last updated on