How We Steer Hermes: an engineering session

Hermes agents run behind chat gateways and on crons, often with nobody watching. When one goes wrong, the tool call has already happened by the time anyone looks at the logs.

We built a plugin that sits inside Hermes and sees every tool call before it runs. For each one it can allow it, deny it, or send the agent an instruction that changes what it does next. Chetan Raghuvanshi (@Chhhee10), founding engineer at Failproof AI, built it, and he'll walk through how it works from the inside.

This is an engineering session, not a product demo. Here's what you'll get into:

Where the plugin hooks into Hermes. Which events we listen to, what the plugin sees at each one, and where in the agent loop a decision can still change the outcome.

Observe, match, decide. How a tool call gets turned into something a policy can check, how policies match on it, and how we pick between allow, deny and instruct.

Keeping it fast. Enforcement used to start a fresh process per tool call and cost around 700ms before a policy even ran. We moved it to a warm daemon and got p50 down to about 0.7ms. Chetan will show what changed and why failing closed matters.

Writing instructions an agent actually follows. Blocking is easy. Getting the agent to do the right thing after a deny is harder, and the wording matters more than you'd think.

What broke along the way. The edge cases, the policies that misfired, and what we'd do differently.

Your questions, live. The last 20 minutes are open. Bring your own Hermes setup and the failure you keep hitting.

Schedule

• 6:00 · how the plugin fits into Hermes

• 6:15 · walkthrough of the code and a live run

• 6:40 · open Q&A

For engineers running Hermes or other agent harnesses who want to see how runtime steering is built. The plugin is open source, so you can follow along in the repo.

Tuesday, October 6 · 6:00 to 7:00 pm · online event