Held. Then released.

PROOF · RECORDED ON OUR OWN ENVIRONMENT · OCTOBER 21ST, 2026

HELD. THEN RELEASED.

Everyone describes the moment. An AI agent is about to do something that needs a person, the action stops, a person decides, the agent carries on. Here is that moment, recorded, 57 seconds, with the clock running.

One action, held in flight. A person said yes, once, from the console, and the agent finished. Every step is on the record. Recorded on our own environment, not a customer's.

What you are watching

A coding agent, Claude Code, is asked to finalize a control assessment in a GRC system. It writes the call and goes to run it. Mountain Theory checks the action first. The rule for this company says finalizing an assessment is a decision a person makes, so the action is held. The agent waits. Nothing runs.

In the console the hold shows up under Held actions with a five-minute clock. Open it and you see why it was held, when it was raised, when it will auto-deny if nobody answers, and six ways to answer: allow or deny, once, for this exact circumstance, or as a general policy. Two of those write a rule.

A person picks "allowed, just this once." The agent runs the call, gets its receipt, and reports back. The record keeps the whole thing: the held action, who answered, what they chose, the result. One row for the auditor.

Why this is the whole product in one minute

Most AI controls work on what the model says. This one works on what the agent does, at the moment it does it. The model still decides. The action does not run until it has been checked. When the rule says a person decides, the person decides, and the agent cannot talk its way past the wait.

HOLD is what makes governed autonomy practical. Turn it on for the narrow set of actions that genuinely need a human, let everything else run at full speed, and keep the record either way.

What we do not claim

This is a recorded run on our own environment, not a customer's. The agent is a real Claude Code session with our hook installed from the console's own installer. The GRC system it calls is a stand-in on the same machine that answers the request; the address on screen is a placeholder. The prompt was typed by a script so the frames could be captured on a timer. The hold, the clock, the answer and every row on the record are the product's, unedited.

The ask

Pick the one action you would least want an agent to take on its own. We will put it on hold in your environment, with your rules and your logins, and show you the record.

Book a demo and bring that action.

Recorded on

Our own environment, October 21st, 2026. One Claude Code session with the Mountain Theory hook, checked against our console. Seven screens from the run, captioned, no edits to what the screens show. The narration and the corner box are Mike May's HeyGen voice and avatar reading the captions. The attribution run from October 7th is here; every published run is on the proof page.

Scroll to Top