Apollo x Tailscale: Introducing “Watcher” for AI Oversight & Control

While AI agents are now powerful enough to do real work, they are too insecure and unreliable to run unsupervised at scale. In production, we see coding agents strategically lying to their users, introducing security vulnerabilities, writing sloppy code, deleting data, or taking initiative far beyond their instructions.

As organizations move from running one experimental agent to a fleet of hundreds, human attention becomes the bottleneck. You cannot manually review every line of code or every API call an agent makes, and this will only get exponentially worse over time. You are forced to make a trade-off between productivity and safety.

Introducing Watcher

We are building Watcher to resolve this trade-off. Apollo Watcher is an oversight layer for AI agents. It detects real-world safety and security failures before they become liabilities, and flags those failures to you. 

It automatically analyzes agent behavior to detect more than twenty distinct failure modes, including instruction violations, deception and dangerous code execution. It is designed for privacy and flexibility: you can run the monitoring backend locally, on-prem, or in the cloud, ensuring sensitive agent logs never leave your control.

If you want to get early access to Watcher, please reach out.

The Apollo x Tailscale Partnership

Today, we are also announcing a strategic partnership with Tailscale to integrate Watcher into their secure LLM gateway, Aperture.

Effective oversight requires two things: access to the data and the ability to understand it.

• Tailscale’s Aperture provides a secure gateway that collects organization-wide agent logs in real-time, handling identity and access without the headache of managing individual API keys.
• Apollo’s Watcher sits on top of Aperture, automatically analyzing every interaction to flag risks before they become incidents.

Key Capabilities

Watcher enables two core workflows for securing your agents.

1. "Analyze" Mode: Organizational Forensics Uncover issues and analyze trends across your entire fleet of agents. Drill down into historical logs to identify recurring failure patterns, audit security risks, or debug why a specific agent went off the rails. Watcher can integrate all deployments across your organization in real-time allowing security experts to see everything in one view. 

2. "Live" Mode: Real-Time Oversight Give individual developers a ‘flight control view’ of their active agent sessions. This enables a single engineer to maintain oversight & control over multiple parallel agents confidently, intervening when Watcher flags unexpected behavior. It can also report the issues straight back to your agent, enabling self-correction.

The Future: From Monitoring to Active Interventions

We are building toward a world where monitors don't just watch–they take actions according to fine-grained guardrail policies. 

• Active Control: Intercepting and blocking harmful actions before they execute.
• Self-Learning Systems: Monitors that, on top of allowing fine-grained customization, can automatically learn your organization’s specific risk profile and preferences.
• Proactive Discovery: Detecting "unknown unknowns" by continuously scanning against a library of thousands of failure modes.

Become a Partner

We are looking for development partners who want to move from unsecured agents to robust, monitored deployments. Please reach out if you’re interested!