Let AI actually run the task

Let AI actually run the task

Most task tools track work. wrrk can do the work.

Every task has an Agent tab. Open it, click one button, and an AI agent reads the task, gathers context from your CRM, drafts a response, and (with your approval) sends it. No tab-switching. No prompt engineering. No copy-paste.

This is the feature that makes wrrk different. Take five minutes to read this article.

Opening the Agent tab

Click any task to open it. Look for the Agent tab, it's the last one, with a robot icon:

Two modes

Before you start a run, pick a mode:

Supervised (default, pick this until you trust the agent) The agent pauses before any action that touches the outside world (sending an email, sending a WhatsApp, changing a deal stage). You see exactly what it wants to do and click Approve or Reject.

Autonomous The agent dispatches low-risk actions (read your CRM, leave comments, create subtasks, search your knowledge base) without pausing. External-messaging actions still require approval, that's a hard rule, not a setting.

Optional: extra context

There's a small text box. If the agent doesn't have everything it needs (a tone preference, a price cap, a specific link), drop it in. Two lines is plenty.

Click "Start agent run"

A live log streams into the panel as the agent works:

You'll see entries like:

• Thinking: the agent's reasoning
• Action: a tool the agent is calling (e.g. get_crm_data)
• Result: what came back

When approval is needed

If the agent decides to send something external, the run pauses and you see this card:

Read what it wants to do. Approve to dispatch. Reject to cancel the entire run.

When it's done

The status flips to Complete and a summary appears at the bottom:

The agent also drops a comment on the task with what it did, so anyone reading the task later has a full audit trail.

What the agent will never do without asking

This is a hard guarantee, not a setting. The agent always pauses for your approval before:

• Sending an email (send_email)
• Sending a WhatsApp message (send_whatsapp)
• Changing a CRM deal stage (update_deal_stage)

If you reject one of these, the entire run is cancelled, the agent doesn't continue without that step.

What it can do without asking

• Read the task and its CRM-linked contact / deal
• Add a comment to the task
• Create a subtask
• Move the task to a different column
• Search your knowledge base
• Fetch a public URL and read its content

What if the agent gets stuck?

It's capped at 8 steps per run. If it can't finish, it leaves a comment explaining what it tried, and the run ends. You can start a new run with extra context to nudge it along.

Cost note

Every agent run uses Claude. The first few runs in a billing period are free, see your plan's usage page for details. We recommend a daily cap (default 50 runs/org/day) until you know your team's pattern.

What's next?

Power-user organization: Subtasks, activity, and time. For admins setting this up safely: Running the AI agent safely.