AI driving AI
Also called: agent orchestration, agent-supervised agents, meta-agent
What does AI driving AI mean?
AI driving AI is when one agent supervises or orchestrates another rather than a person doing it. The supervising agent watches a run, answers the decisions it stops on, sends instructions and interrupts, using the same interfaces a human supervisor would reach for.
What has to exist for this to be possible.
It only works if the supervision surface is programmable, not just visual.
The run has to be legible
A supervising agent needs the session as structured data: what is happening, what is pending, what is blocked. A screen a human reads is not enough, because there is nothing for another agent to reason over.
Decisions have to be addressable
A pending permission request needs to be a thing with an identity that can be answered by a call. If approving is only a button, only a person can approve.
Authority has to be attributable
If an agent approved something, the record has to say so. Otherwise you have autonomy with no way to reconstruct who or what authorised an action.
Why anyone would want this.
Mostly because the alternative is a person answering routine questions all night.
Routine decisions do not need you
Plenty of what an agent stops to ask is unremarkable. A supervising agent can clear those and escalate only the ones that need judgement.
Fleets need coordination
Several concurrent runs are more supervision than one person can hold. Orchestration is how that stays manageable.
It raises the stakes on the record
Handing approval to an agent is exactly when an attributable audit trail stops being a nice-to-have.
Commonly confused with.
Three things this is not.
AI driving AI questions
What does AI driving AI mean?
One agent supervising or orchestrating another rather than a person doing it: watching the run, answering the decisions it stops on, sending instructions, and interrupting, through the same interfaces a human supervisor would use.
What does Virex provide for this?
Virex Remote exposes the session and decision surface over MCP, so an orchestrating agent can follow a run and act on it the same way a person can from the browser.
Does this remove the human?
It should not. The sound pattern is an agent clearing routine decisions and escalating the consequential ones, which is why attribution in the record matters more, not less.
How is it different from subagents?
Subagents are structure inside one run. This is one agent supervising a separate run from outside, with no shared session.
Is the approving agent recorded?
It needs to be. If an agent approved an action, the audit trail has to say that rather than implying a person did.
A supervision surface both can use.
Virex Remote exposes sessions and decisions over MCP as well as in the browser.