Human-in-the-Loop (HITL)

HITL is the principle that no agent output reaches a client or an external system without a human making an explicit decision first. It is not an optional layer — it is enforced at the database and tool-dispatcher levels so that no code path can bypass it.

Technical enforcement: Governance & Guardrails — Layer 3 HITL spec, approval DB schema, resolution flow code Approval display: UI Patterns — how approvals look in each app Full approval lifecycle: Approval Flow — creation → notification → decision → downstream

The Non-Negotiable Rule

No agent can trigger a write action on an external system
without a corresponding approvals record with status = 'approved'.

This is enforced at the tool dispatcher — before any external API call executes, it checks for an approved record. The UI cannot bypass it. An agent cannot bypass it. Even a super admin calling the API directly cannot bypass it.

Where HITL Appears

HITL gates exist at every point where the pipeline crosses a boundary — from research to content, from content to publish, from data to spend.

1. Onboarding Chain

Nothing in the content pipeline runs until the strategy and deliverable plan are approved. These are the upstream gates for everything else. The Action Item Planner runs in parallel with the deliverable plan approval wait — its tasks appear for DM review while the plan itself may still be pending.

Autonomy settings: All four onboarding gates (context, strategy, deliverable plan, action items) can be bypassed per tenant or platform-wide via System → Autonomy. When disabled, the pipeline auto-approves the output and immediately triggers the next stage. The dashboard shows an “Auto-approved” badge in violet and the timeline records a System entry for full auditability. See Autonomy Settings.

2. Content Pipeline

3. SEO Pipeline

4. Paid Ads Pipeline

Ad-related approvals are always risk_level: high — they have immediate spend impact.

5. Reporting & Analysis

6. Agent Mid-Run Decisions

Some agents encounter a decision point mid-execution that they cannot resolve themselves. They create an approval gate and pause, waiting for a human to decide before continuing.

7. System Safety Gates

Approval Types Summary

Risk levels by deliverable (content_review only):

Who Reviews What

Some approvals require both a DM reviewer and a tenant admin sign-off (notably email campaigns and ad optimization pushes). Both approvals must be resolved before the gate opens.

Related Docs