New FREE Hermes AI Mission Control Update!

The video presents Hermes Labyrinth, a newly released free plugin for the Hermes AI agent (developed by Noose Research), designed to solve the 'black box' problem that plagues AI agent workflows. The presenter, a digital avatar of Julian Goldie (CEO of Goldie Agency), opens by identifying a core frustration: when you send an AI agent off to do business tasks like content research, competitor analysis, or email handling, you receive only the final result with no visibility into what happened in between — what tools were used, what failed, what models were switched, or what decisions were made.

Labyrinth addresses this by acting like a flight data recorder for AI agents. It converts every action the agent takes into what it calls a 'crossing' — a discrete, inspectable step in the agent's journey. These crossings capture prompts, tool calls, tool results, failures, model switches, sub-agent spawns, approval requests, memory retrievals, secret redactions, and context compressions.

The presenter walks through eight key components of the plugin. The Journey Index serves as a home base showing all recent agent tasks from any source — command line, dashboard, chat gateway, cron jobs, or delegated tasks. The Labyrinth Map provides a visual, chronological layout of every step an agent took, including dead ends and shortcuts, helping business owners identify inefficiencies. The Inspector allows users to click into any individual step to see full input/output data, duration, success/failure status, and supporting evidence — useful for debugging or client reporting. Gypose auto-generates observations based on evidence in the data, surfacing important findings without requiring users to manually review every line. The Skill Atlas displays all skills the agent has built over time — bundled, optional, external, and custom — giving users visibility into the agent's accumulated knowledge base. The Chronate tracks scheduled automations, showing next run times, last failure points, and execution sources. The Model Ferry logs every instance where the agent switched between AI models, which the presenter highlights as especially valuable for cost management, as it reveals exactly where expensive model usage is occurring. Finally, the Reports feature allows users to export any journey as a clean markdown or JSON file with secrets automatically redacted, making it safe to share with clients or team members.

A major design emphasis is that Labyrinth is strictly read-only — it can observe everything but cannot start, stop, modify, or create any agent sessions. The presenter frames this as a critical safety feature, arguing that most tools requiring deep agent access introduce risk of unintended interference. Labyrinth also automatically redacts API keys and passwords from all previews and reports.

Setup is described as minimal: two commands to create a plugins folder, clone the Labyrinth repo, and restart the Hermes dashboard. The presenter estimates setup takes less time than making a coffee. While the initial setup requires basic command-line comfort, the ongoing interface is a visual dashboard requiring no coding or configuration.

The presenter concludes by arguing that Labyrinth represents a turning point for AI agent adoption in real business contexts — that the inability to see what agents were doing was the primary barrier to trusting them with larger tasks, and that Labyrinth removes that barrier, enabling businesses to scale agent usage more confidently.