Delegating to a sub-agent

A sequence of actions on an interface (look, act, re-read, start over) quickly fills an agent's context with captures. It is better to hand it to a sub-agent, which returns only the result.

What you ask

"Read the version number shown in the application's About window."

What the agent does

1. It calls vue_pilot_brief with the objective. The tool uses no AI: it returns a go/no-go gate and a ready-to-paste prompt for a sub-agent.
2. If the gate is ready, it launches the sub-agent with this prompt. The captures stay inside the sub-agent; only the extracted text comes back up.
3. If the gate is not ready, it calibrates first with vue_calibrate, which freezes the screen topology under a name, rather than driving blind.

Two tools complement the brief:

• vue_profile says whether the current screen topology matches a known profile;
• vue_profile_brief produces a sub-agent's starter: the profile's matching gate and addressable screen indices.

What you get

The task's result, without the pile of intermediate captures.

What does not change

The sub-agent works under the same ceiling as the main agent: it cannot do more. Exclusions, censoring and the audit log apply to it like to any client.

Next: When it doesn't work.