Vigilator

Vigilator is a human-in-the-loop supervision platform for AI agents, built by Coloop LTD. It sits between autonomous agents and high-stakes actions, intercepting decisions that require human review and routing them to the right person before the agent proceeds. The platform is framework-agnostic and integrates via a plain HTTP API with an OpenAPI spec.

What It Is

Vigilator is a control layer designed for teams running AI agents in production. When an agent encounters a decision it cannot make alone—or one that must be approved—it raises an "interrupt" that lands in a shared inbox. Reviewers can approve, reject, or edit the agent's proposed tool call arguments, with the full conversation context visible alongside the request. The platform also provides a live view of every connected agent, analytics on approval rates and response times, and a searchable audit log of all actions taken by both humans and agents.

How the Interrupt Workflow Operates

The core workflow centers on interrupts: structured pause points where an agent hands a decision to a human. Each interrupt surfaces the full conversation history and the exact tool calls the agent wants to run. Reviewers act from a shared inbox, and if no one responds within a configured time window, escalation chains automatically hand the request to the next reviewer. Notifications arrive by email or in-app, and decisions can be pushed into external infrastructure via webhooks.

Supervision and Observability Features

Beyond the inbox, Vigilator offers a Live View that shows what every connected agent is doing in real time, enabling teams to spot unexpected behavior and intervene before it causes harm. Interrupts and sessions can be tagged and classified, then routed to the team that owns that decision type. An optional AI co-worker called Argus can summarize and classify each interrupt, flag risky requests, and draft a suggested decision to speed up reviewer response time.

Access Control, Roles, and Audit

Organizations on Vigilator are fully isolated from one another. Within an organization, fine-grained roles and teams control who can view, decide, and escalate. Every action—approvals, rejections, escalations, reassignments, role changes, API key creation, and webhook edits—is recorded in a searchable audit log that can be filtered by person, category, and time, and exported to CSV.

Integration Model

Vigilator exposes a plain HTTP API with an OpenAPI specification and an interactive reference, meaning any agent framework that can make an HTTP request can raise an interrupt. The platform does not require a specific SDK or agent framework, making it compatible with custom-built agents and popular frameworks alike.