Skip to Content
Welcome to 10ex Docs — explore Tutorials, Guides, Reference, Concepts, Use Cases, and the Agent Store.
ConceptsApproval workflows
Concept

Approval workflows

Some agent outputs are reversible. A draft blog, a market research report, a creative brief. Others aren’t: a cold email already in someone’s inbox, an ad that just spent budget. 10ex separates those two classes through approval workflows. Any agent run that produces an irreversible artifact can require explicit human approval before the irreversible step happens.

Active Agents page with one agent showing Approval Required badge and inline Approve and Reject buttons
Active Agents, Marcus/Blog awaiting approval. Approve or Reject inline; the run pauses until you act.

The two-stage shape

Every approval-gated agent runs in two stages.

  1. Generate. The crew runs end-to-end and produces a candidate artifact, but stops before the side-effect.
  2. Send, publish, or spend. This happens only after a human clicks Approve (or an MCP approve_* tool fires).

If a human clicks Reject, the run closes without the side-effect. The artifact is preserved in run history; nothing reaches the outside world.

Which agents gate by default

PersonaCrewDefault gateWhy
Novaemail-sequence-generationRequiredSending cold emails is irreversible and reflects on your sender domain
Orionemail-outreach-senderConfigurable per runSingle-shot sending; trusted senders may turn it off
Marcus/Blogblog-generationRequired for publishDraft is free; pushing to Webflow/Sanity gates
Zara / Ravi / Elena*-ads-managerRequiredLive campaigns spend real money
Atlaswebinar-generatorOptionalAvatar video generation is reversible until the calendar invite goes out

Agents whose output is purely informational (Sora, Juno, Maya, Marcus/Prospector) don’t gate. They just deliver.

How it shows up in the product

Active Agents (/agent-hub/active-agents) is the operational hub. Each tile shows status:

  • Ready to run. No pending action.
  • Approval Required. A yellow banner with Approve and Reject buttons appears below the tile.
  • Failed. The run errored; the tile shows the failure reason.

When a tile is in Approval Required, the Configure Agent button is replaced with the inline approve/reject pair. Clicking Approve continues the original run with the side-effect step. Clicking Reject closes it.

How it shows up over MCP

The same gate is exposed as paired tools. Every gated agent has a preview_* and an approve_*.

MCP prompt

preview_email_sequence({ segment_id, brand_id, objective: "book a demo", sender_connector_id, num_emails: 3 }) // returns { preview_id, recipients, sample_emails: [...] }

MCP prompt

approve_email_sequence({ preview_id }) // only this call actually starts sending

MCP prompt

launch_email_sequence({ ..., approval_required: false }) // shorthand to skip the gate when you trust the input, e.g. an internal nurture against your own list

The pattern repeats: generate_blog then publish_blog, generate_ad_copy then launch_ad_campaign, generate_webinar then calendar invite.

Why approvals matter

The asymmetry is the whole point. A bad draft costs you a minute of reading. A bad send costs you a domain-reputation hit, an unsubscribe wave, or a six-figure ad spend on the wrong creative. Approvals trade a few seconds of friction for a kill switch on the actions that can actually hurt you.

A common misunderstanding: teams disable approvals to “speed things up,” then ship a typo to 10,000 prospects. The fix is rarely faster automation. It’s a cheaper preview and a clearer review surface, which is what the gate gives you.

Audit trail

Every approve or reject action writes a row to the mcp_tool_call audit table (for MCP-driven approvals) or the equivalent in-app audit log (for UI clicks). Each row stores:

  • Who approved (user_id) or what (api_key prefix)
  • When (timestamp)
  • Which run (preview_id / sequence_id)
  • The approved payload’s hash, so a reviewer can prove the exact artifact that was approved is what shipped

This makes approval workflows the natural fit for compliance regimes: GDPR for outbound, ad-platform policy review, Webflow editorial sign-off.

Designing for the gate

Two practical rules.

  1. Preview should be cheap. Cost gates produce friction; if previewing burns 10 credits, people skip the preview. 10ex’s pricing reflects this. Preview is free for Nova, and the per-email meter starts on send.
  2. Approve should be batched. When you approve a 50-recipient sequence, you approve the plan, not 50 individual emails. The reviewer sees a sample (typically 3 emails) and the sequence cadence; the rest is generated against the same template.

Bypassing the gate

Some teams operate “fast lane” by sending against pre-vetted templates against owned lists. Two ways to bypass:

  • At hire time. Uncheck Approval Required in the agent’s Configure panel.
  • At run time. Pass approval_required: false in the MCP launch_* call.

Both bypass options write a clear audit entry, so a reviewer can spot lanes that have been disabled.

Common questions

Who can approve? Any workspace member can approve UI-side. MCP approvals are scoped to the API key’s role. There’s no separate “approver” role today; that’s tracked as tech debt.

What happens if I forget to approve? The run sits idle. There’s no auto-expire in v1. Reject explicitly to clean up.

Can I approve from Slack or email? Not in v1. Approvals are UI or MCP only. Slack triggers are on the roadmap.

Last updated on
Launchpad: strategic playbookFlows