Introduction

The inbox is a section of the Vigilator dashboard that allows humans to manually classify and handle problems encountered by AI agents.

When an agent encounters a problem an "interrupt" is sent to the inbox. A human can then manually handle the problem and govern the next steps.

Example use cases

Here are some examples of use cases where you may want to make use of the inbox feature.

Irreversible or High-Risk Actions

• Financial Transactions - An agent needs to refund a customer, make sure a human has the final say.
• Deletion of production data or specific infrastructure - Deleting user accounts
• Sending external communication on behalf of a company - Emails to customers, Public Posts, Legal Correspondence, etc

Legal, compliance, and regulatory exposure

• Content moderation decisions that could result in banning real users or accounts
• Modifying pricing, discounts, or contract terms with customers
• Anything affecting hiring, firing, promotion, or performance decisions

High Ambiguity or Low Model Confidence

• When the agent's own confidence score is low, or when it's operating outside the distribution of tasks it was tuned for
• Novel situations not covered by existing policy/playbooks
• Conflicting or contradictory instructions in context

Safety Critical or High Stake Environments

• Anything that could physically affect people (robotics, autonomous vehicles, medical devices)
• Content that could be defamatory, discriminatory, or harmful if published
• Security-sensitive actions (granting permissions, rotating credentials, modifying access controls)

Industrial & Physical Systems

• Actions in shared human-robot workspaces
• Force/torque thresholds being exceeded
• Path planning through ambiguous or contested space
• Loss of sensor confidence

As demonstrated, there are a wide range of scenarios where HITL may be critical to the integrity of your business operations. Vigilator provides you with the platform that allows your employees ("humans") to make critical decisions.

Technical implementation

Ready to wire an agent up? The Technical section covers the API and SDKs:

• Python SDK - open interrupts, learn their outcome, and stream Live View sessions.
• Webhooks - resume a paused run the instant its interrupt is answered.
• API keys - authenticate your agents.

Framework-specific guides (LangGraph, LangChain, DeepAgents, Mastra and TypeScript) are on their way.

Working the inbox

• Navigating the inbox - the list, scopes, sorting and live updates.
• Reviewing an interrupt - deciding action requests, triaging, Argus and the audit log.
• Manager review - the escalation queue.
• Shortcuts & command palette - work the inbox from the keyboard.