Agent Tasks: delegate multi-document work
Use Tasks to run an unattended agent that plans, estimates credits, and delivers documents. Credits vs chat, ADP/Mistral, and launch limits.
Agent Tasks let you delegate a bounded GRC document job instead of chatting turn by turn. The agent proposes a short plan with a credit estimate, waits for your approval, then writes the deliverables offline.
Open Tasks in the app sidebar (or switch to Agent where the Chat | Agent toggle appears).
How a task runs
- Describe the job (for example: five core policies for a remote German tech startup).
- The agent returns a plan: list of deliverables and a credit band (1, 2, or 4 credits).
- Approve (or reject). Credits are reserved only when you approve.
- The agent writes each deliverable, then packages results as markdown, a zip bundle, and office formats when conversion is available.
- Download from the task when status is completed. You can cancel a running task; reserved credits are refunded.
Agent credits vs chat limits
These are separate.
| Chat | Agent Tasks | |
|---|---|---|
| What is limited | Messages and model usage in a 4-hour window | Task credits (runs) |
| Reset | Every 4 hours (UTC bins) | Paid plans: calendar month. Free: 3 credits lifetime (not monthly) |
| When it spends | Each chat / Beyond turn | When you approve the plan |
How many task credits do you get?
- Free: 3 lifetime
- Essential: 50 / month
- Plus: 125 / month
- Standard: 250 / month
- Pro: 500 / month
- Business: 1000 / month
Each approved task costs 1, 2, or 4 credits depending on how many deliverables the plan lists.
Models and Advanced Data Protection (ADP)
- Default (ADP off): Agent planning and writing use the Grok model path (aligned with the default paid chat assistant).
- ADP on (your account setting, or your workspace organization ADP): Agent planning and writing use Mistral (EU), same family as chat under ADP. Grok/xAI is not used for those model calls.
- You do not pick a model per task today; the worker chooses from your ADP posture.
Multi-document packs
When a task produces several documents, later documents follow the same section outline as the first document so the pack stays consistent. Content still differs per document.
Not included at launch (roadmap)
- Web research inside agent planning
- Skill packs and full workspace URL / instruction parity with chat
- Connectors (for example Notion read/write from the agent)
Tips
- Name the company, country, frameworks, and audience in the prompt.
- Prefer a clear deliverable list over a vague full ISMS.
- Approve the plan only when the deliverable list and credit band match what you want.
Related
- In-app product changelog (User menu)
- Account and billing