Skip to main content

AI Agent

FlowMail's AI agent lets you give natural language instructions through the command palette (Cmd+K). Instead of navigating menus, you describe what you want and the agent figures out the steps.

How It Works​

Open the command palette with Cmd+K, switch to agent mode, and type what you want:

  • "Draft a reply to Sarah's email about the budget review"
  • "Create a plan for preparing next week's board presentation"
  • "Search my emails for anything about the Henderson contract"
  • "Create a new Card for the website redesign project"

The agent breaks your instruction into steps, uses the appropriate tools, and streams its progress back to you in real time. When it's done, the result appears as a plan in the Plans view with a full transcript of what it did and why.

What the Agent Can Do​

The agent has access to a curated set of tools:

  • Email — Search, read, and compose emails
  • Plans — Create plans from natural language instructions
  • Cards — Create and manage context cards
  • Drafting — Generate email drafts with full context
  • Search — Semantic search across your knowledge base using vector embeddings
  • Calendar — Create and check calendar events
  • Knowledge — Look up entities and relationships

Each tool call is visible in the agent's execution transcript, so you can always see exactly what it did.

Progressive Trust​

Not every agent action happens automatically. FlowMail uses a progressive trust system:

  • Ask every time — The default for sensitive actions like sending emails. The agent pauses and shows you what it wants to do.
  • Always allow — For actions you've approved before (like searching emails), you can grant standing permission so the agent doesn't need to ask again.

Trust is scoped — approving "search emails" doesn't approve "send emails." You stay in control of what the agent can do autonomously.

Approval Flow​

When the agent needs your approval for an action, it creates an approval item in your Focus Queue. You see exactly what the agent wants to do — the tool it wants to call, the arguments it will use — and you can approve, reject, or modify it before it proceeds.

This means the agent never takes sensitive actions without your explicit sign-off. You can queue up instructions and review the results at your own pace.

Agent Transcript​

Every agent execution is recorded as a plan with a full transcript. Open the Plans page to see:

  • Each step the agent took
  • Which tools it called and with what arguments
  • Its reasoning at each decision point
  • The final outcome

This transparency means you can always understand why the agent did what it did, and provide feedback if it got something wrong.

Context Enrichment​

The agent doesn't work in a vacuum. When you give it an instruction, it automatically gathers relevant context:

  • The full email body and thread history (when working with a specific email)
  • Knowledge about the sender from your knowledge base
  • Card instructions and context for the relevant card
  • Your learned rules and preferences

This context helps the agent make better decisions without you having to spell out every detail.