Agents that pass an exam

The creation chain, one step at a time

Today one small, older model writes a new agent in a single pass. The new chain splits that work into steps, each with its own brief and its own model, and each step checks the one before it.

• BUILTIntake: the owner says in plain words what the agent is for and who it serves.Evidence: Its step brief is on main (resources/agent-creator/v2/steps/intake.md) and runs through the step runner (app/Services/AgentBuilder/CreatorV2/CreatorV2Runner.php); nothing in the Studio calls the runner yet.
• BUILTExpert map: what an expert in this field knows, asks and refuses.Evidence: Its step brief is on main (resources/agent-creator/v2/steps/expertise.md) and runs through the step runner (app/Services/AgentBuilder/CreatorV2/CreatorV2Runner.php); nothing in the Studio calls the runner yet.
• BUILTCharacter: temperament, values and the way the agent speaks.Evidence: Its step brief is on main (resources/agent-creator/v2/steps/character.md) and runs through the step runner (app/Services/AgentBuilder/CreatorV2/CreatorV2Runner.php); nothing in the Studio calls the runner yet.
• BUILTPersona: the agent's own instructions, written from the steps before.Evidence: Its step brief is on main (resources/agent-creator/v2/steps/forge.md) and runs through the step runner (app/Services/AgentBuilder/CreatorV2/CreatorV2Runner.php); nothing in the Studio calls the runner yet.
• BUILTGuard: what the agent will not do, and when it hands a visitor to a person.Evidence: Its step brief is on main (resources/agent-creator/v2/steps/guard.md) and runs through the step runner (app/Services/AgentBuilder/CreatorV2/CreatorV2Runner.php); nothing in the Studio calls the runner yet.
• BUILTKnowledge: the facts the agent needs from the owner's own material.Evidence: Its step brief is on main (resources/agent-creator/v2/steps/knowledge.md) and runs through the step runner (app/Services/AgentBuilder/CreatorV2/CreatorV2Runner.php); nothing in the Studio calls the runner yet.
• BUILTVoice: how the agent sounds when it speaks.Evidence: Its step brief is on main (resources/agent-creator/v2/steps/voice.md) and runs through the step runner (app/Services/AgentBuilder/CreatorV2/CreatorV2Runner.php); nothing in the Studio calls the runner yet.
• BUILTFace and eight looks: the brief for the agent's face and its eight looks.Evidence: The face and looks briefs are on main (resources/agent-creator/v2/steps/face_brief.md and resources/agent-creator/v2/steps/looks_brief.md) and run through the step runner; nothing in the Studio calls the runner yet.
• NEXTHarness and house: the tools, memory and house rules the agent works inside.Evidence: Not built: no harness and house step is on main.
• BUILTExam: the agent is graded in its own field before it can sell.Evidence: The exam run that asks, grades, repairs once and retests is on main (app/Services/AgentBuilder/CreatorV2/Exam/ExamRunner.php), behind ai.features.creator_v2_exam with the creator key; no agent has sat it for sale yet.
• BUILTCopy: the words the agent's page uses to describe it.Evidence: Its step brief is on main (resources/agent-creator/v2/steps/copy.md) and runs through the step runner (app/Services/AgentBuilder/CreatorV2/CreatorV2Runner.php); nothing in the Studio calls the runner yet.

The field exam

Nothing tests today whether a new agent answers like someone who knows the field. The exam closes that gap.

• BUILTTen questions in the agent's own field, written and graded by a model from a different family than the one that wrote the agent, plus one safety probe. A score under seventy-five or any safety fail means one repair and a retest; a second miss leaves the agent a draft that cannot sell.Evidence: The exam writer, grader and repair briefs are on main (resources/agent-creator/v2/steps/exam_writer.md, resources/agent-creator/v2/steps/grader.md, resources/agent-creator/v2/steps/repair.md). The exam run that asks, grades, repairs once and retests is on main (app/Services/AgentBuilder/CreatorV2/Exam/ExamRunner.php), behind ai.features.creator_v2_exam with the creator key; no agent has sat it for sale yet.

Regulated fields get guardrails

Law, medicine, money, mental health and licensed trades need more than a good prompt. An agent in one of these fields carries a fixed house block that the model never writes and cannot change.

The block says: I'm an AI with general information, not a licensed professional. If this is an emergency, contact your local emergency services first. I can hand you to a person who can help with your own situation.

No NetShow agent claims a license or a credential, and none is described as a licensed professional.

• BUILTA fixed house block for regulated fields: an AI with general information, emergencies first, and a person to hand off to.Evidence: The fixed block the model never writes is on main (app/Services/AgentBuilder/RegulatedFieldGuardrails.php), added at birth behind ai.features.creator_regulated_guard when the classifier (app/Services/AgentBuilder/CreatorV2/RegulatedFieldClassifier.php) names a regulated field.

A roster of ready employees

The same chain makes category experts ahead of time, so an owner can hire one that has already passed its exam instead of building from a blank page.

• NEXTCategory experts born through the chain and sold on the AI employee ladder, where Front Desk starts at $699 a month and the exact price is set in writing after a build call.Evidence: The ladder and its prices are in config/ai_employees.php; no roster agent has been born through the chain yet.

Talk with us

This page stays open at the address in the code on the first slide.

NetShow (949) 828-2040 support@netshow.ai Laguna Niguel, Orange County, California, USA