Patterns
9 design moves that make an agent safe to use, grouped into four stages of the agent’s work.
Stage 1
Before it acts#
Get a clear, informed yes before anything changes.
- Show the exact change before it happensThe person approves the exact change; the agent runs only that.Seen in Notion Agent, Zapier Agents & AI by Zapier, Intercom Fin, Shopify Sidekick
- Put the review where the evidence isThe person approves the change on the field it changes.Seen in Shopify Sidekick, Figma Design agent, Microsoft Excel agents, Linear Agent
Stage 2
While it works#
Keep permission and progress clear while unattended.
- Let the agent act alone only within written rulesInside the rule the agent acts; outside, the agent asks the owner.Seen in Atlassian Rovo, Asana AI Teammates, Notion Agent, Airtable Omni & field agents
- Keep working after the tab closesThe job runs in the product, not in the browser tab.Seen in Shopify Sidekick, Linear Agent, Zapier Agents & AI by Zapier, Notion Agent
- Give every agent job one owner and one rulebookEvery door into the product runs the same job, same rules.Seen in Agentforce Service, HubSpot Customer Agent, Asana AI Teammates, Shopify Sidekick
Stage 3
After it acts#
Prove what happened; offer an honest way back.
- Be honest about what undo can undoUndo restores what it safely can and names what it can’t.Seen in Microsoft Excel agents, Figma Design agent, Intercom Fin, HubSpot Customer Agent
- Leave a receipt, not just a chat messageAsked, approved, changed, failed: all on the record.Seen in Notion Agent, Agentforce Service, Linear Agent, Airtable Omni & field agents
Stage 4
Is it working?#
Measure real results; decide what to build.
- Measure cost per result that actually workedDivide spend by checked results, not by attempts.Seen in Linear Agent, Intercom Fin, HubSpot Customer Agent, Agentforce Service
- Turn every bad run into a fix someone ownsA bad run becomes an owned ticket and a test case.Seen in Intercom Fin, HubSpot Customer Agent, Atlassian Rovo, Agentforce Service