Agent Curriculum System

Section 01

What is this?

This is a curriculum builder for your AI agents. Think of it like a training program — the kind a manager would write before onboarding a new employee. You define what each agent needs to know. The tool turns that into instructions the agent can actually follow.

Right now, your agents do their jobs based on their system prompts. But if a skill isn't defined clearly — with a real example, a specific task, and a clear picture of what "correct" looks like — the agent fills in the blanks on its own. Sometimes that works. More often, it loops, guesses, or stops at the wrong time.

This tool fixes that. You write the curriculum once. Your agents work from it every time.

The core idea

An AI agent is not intelligent — it is a mirror. It reflects back what you give it. If you give it a clear curriculum, it teaches clearly. If you give it nothing, it invents something, and that invention is usually wrong.

Section 02

What problem does it solve?

If you have ever told an agent to "teach the other agents tasks" and watched it either stop randomly or repeat the same lesson over and over — this is the fix.

That problem happens because the teaching agent has no answer to three questions:

1

What am I teaching?

Without a defined domain, the agent picks topics at random. It might teach the same thing twice, or teach something completely unrelated to the agent's actual job.

2

How do I know if the student got it?

Without a definition of "correct output," the agent has no way to evaluate. It either always passes, always fails, or loops forever trying to decide.

3

When do I stop?

Without a stopping condition, the agent keeps going. More lessons, more repetition, no end point. The loop you experienced was this exact problem.

This tool gives the agent answers to all three questions before it starts. That is what a curriculum is — a pre-defined set of answers to those questions.

Section 03

How does it fit into your factory?

Your SMS Software Factory already has a pipeline. Tickets come in, move through stages, and agents do their work at each step.

Ticket

→

PM

→

Architect

→

Task

→

Builder

→

Review

→

Test

→

Deploy

This tool does not add anything new to that pipeline. It works through what you already have.

Here is how: one of the three outputs this tool generates is called a Training Ticket. A training ticket looks exactly like a regular ticket — because it is one. You submit it into your factory the same way you submit any other ticket. The difference is the content: instead of asking your agent to build something, it asks your agent to demonstrate a skill.

Your factory processes the training ticket. Your agent responds. The teacher evaluates the response against the expected output you defined. If the agent passes, it moves to the next skill. If it fails, it tries again. When it has hit its attempt limit without passing, it stops and flags the skill for your review instead of looping.

No new infrastructure

You do not need to build a separate training system. Your pipeline is the training system. This tool just gives it the right curriculum to work from.

Section 04

Your agents and pre-built skills

The tool comes pre-loaded with your factory agents and a starter curriculum for each — including a Pipeline Agent aligned with Gauntlet Night School (managed agents, orchestration, guardrails).

Agent	Role in your pipeline	Pre-built skills
Router	Classifies and routes incoming tickets before they enter the pipeline	3 skills
PM Agent	Turns tickets into structured Spec documents	3 skills
Coding Agent	Builds from task lists; reviews its own output against the Spec	3 skills
Sales Agent	Converts client requests into scoped proposals	3 skills
Pipeline Agent	Orchestrates all agents — task decomposition, assignments, guardrails, stop conditions (Night School)	6 skills
Custom Agent	Define your own agent if it is not listed above	Start fresh

Section 04b

Gauntlet Night School → your factory

This curriculum is aligned with Gauntlet AI Night School (managed computer agents & software factories). The lecture’s mental model maps directly to SMS Software Factory:

Night School concept	Your factory
Managed agent = model + tools + governed environment	VPS at factory.smesolutions.energy, auth, audit log, Ollama
“What is it allowed to touch?”	Router scope validation + Pipeline guardrail gates skill
Orchestration / Kanban dispatch	Pipeline Agent → `tickets/agent-task-queue.json` + learning cycle
Software factory team (Kelly Claude)	PM → Architect → Task → Builder → Review → Test → Deploy
Model routing (planner vs coder)	Ollama on VPS; frontier planning vs local execution (configurable)
Human approval before ship/deploy	Deploy stage + Pipeline “Guardrail Gates” skill
Four failure modes (drift, hallucination, ambiguity, permissions)	Pipeline “Failure Mode Detection” + curriculum stop/flag rules
Teacher curriculum with expected output	This tool — fixes agents looping or inventing lessons

Start here for Pipeline Agent

Open Agent Curriculum → select Pipeline Agent → review the six pre-built skills → Save to Factory → run a learning cycle from the dashboard.

Section 05

How to use it — step by step

The tool has three screens. You move through them in order: Select Agent → Build → Output.

1

Select Agent

Click the agent you want to train. The tool loads a starter curriculum for that agent automatically. You will see where it sits in your pipeline and what it is responsible for producing.

2

Build the Curriculum

Each skill has five fields to fill in. You click through skills at the top and edit each one. The fields are:

Skill Name — What you are calling this skill. Keep it short.
Concept — What this skill is and why it matters. One or two sentences.
Example — A real, specific demonstration. Not vague. Something your agent could actually look at and understand.
Task — What you ask the agent to do. This is the assignment.
Expected Output — What the correct response looks like. This is the most important field. Be specific. The agent is evaluated against this.

You also set Max Attempts — how many tries the agent gets before the skill is flagged for your review. The default is 3.

3

Generate and Copy Output

Click "Generate Output." You now have three outputs ready to copy. Each one has a copy button. See Section 06 for what each output is and exactly where it goes. You can also click Save to Factory to write all outputs to disk and queue training tasks automatically.

The most important field

Expected Output is the field most people skip because it is the hardest to write. Do not skip it. If you cannot describe what correct looks like, the agent cannot evaluate itself. Start simple — even one sentence is enough to begin.

Section 06

Your three outputs

Every curriculum you build generates three separate outputs. Each one goes to a different place in your factory.

Output 1

Teacher Prompt

This is the system prompt for a teaching agent. It contains the full curriculum — all skills in sequence, the lesson flow, the student state tracker, and the required output format. The teaching agent reads this and knows exactly what to teach, how to evaluate, and when to stop.

Where it goes: Paste this into the system prompt of your teacher agent. If you do not have a separate teacher agent yet, this prompt is what you use to create one. Or click Save to Factory — it writes to skills/curriculum/{agent}-teacher.md.

Output 2

Training Tickets

One ticket per skill. Each ticket is structured to exercise that specific skill through your normal factory pipeline. You submit these into your factory the same way you submit any other ticket. The agent processes it, produces output, and the teacher evaluates it.

Where it goes: Submit these into your factory's Ticket Input one at a time, in order. Or click Save to Factory — they are written to tickets/training/ and queued for the agent learning cycle.

Output 3

Memory & Audit Format

Structured JSON formats for recording skill progress. After each lesson, these entries go into your factory's Memory Log and Audit Log. This is how the learning persists — so your agent remembers what it has mastered and you can track the training history over time.

Where it goes: Use these formats when writing to your existing Memory Log and Audit Log. They are designed to match the structure your factory already expects. Save to Factory writes them to skills/curriculum/{agent}-memory-format.md.

Section 07

Quick start — fastest path

If you want to see this working as fast as possible, follow these five steps:

5-Minute Quick Start

Step 1

Open the tool at /curriculum. Click PM Agent.

Step 2

Go to Build. Look at Skill 1 — "Requirements Extraction." Edit the Expected Output field to match what your PM Agent actually produces right now.

Step 3

Click Generate Output. Copy the Teacher Prompt — or click Save to Factory.

Step 4

Paste the Teacher Prompt into a new agent's system prompt.

Step 5

Copy Training Ticket 1. Submit it into your factory. Watch what happens.

You do not need a perfect curriculum to start. One skill with a clear expected output is enough to see the difference.

Section 08

Common questions

Do I need to use all the pre-built skills?

No. Delete any skill that doesn't apply to how your agent actually works. The pre-built skills are a starting point, not a requirement. Your agent's real behavior is what matters — build the curriculum around that.

What if my agent fails every skill?

That is useful information. It usually means the Expected Output you defined does not match what your agent is capable of producing yet. Start by making the Expected Output simpler, or go back to the agent's own system prompt and check whether the instructions there are clear enough for the task.

What does "flagged" mean?

Flagged means the agent used all of its attempts on a skill without passing. The teaching agent stops, writes a flag entry to your Audit Log, and waits. It does not loop. You review the flag, adjust the skill definition or the agent's system prompt, and run the training ticket again.

Does this replace my agents' existing system prompts?

No. The Teacher Prompt goes into a separate teaching agent. Your existing agents keep their existing system prompts. Training happens alongside normal operation — training tickets go through the same pipeline as regular tickets.

What if I want to train an agent that isn't listed?

Select Custom Agent. Fill in what the agent receives, what it produces, and define your skills from scratch. Everything else works the same way.

How often should I run training?

Run it whenever an agent starts producing unexpected output, when you change an agent's system prompt, or when you add a new type of ticket to your factory. Think of it like re-onboarding — any time the job changes, the training should follow.

Section 09

Key terms

Curriculum

The full set of skills an agent is being trained on, in order. Defined by you. The agent cannot change it or add to it on its own.

Expected Output

Your description of what a correct agent response looks like for a given task. The teaching agent compares the student's actual response to this. If they match, the skill passes.

Training Ticket

A ticket submitted into your factory specifically to teach and evaluate one skill. Processed by your normal pipeline. Not a production ticket — it is labeled as training so your agents know the context.

Student State

A running record of which skills have been mastered, which are flagged, which skill is currently active, and how many attempts have been used. The teaching agent updates this after every lesson.

Max Attempts

The number of tries an agent gets on a skill before the teaching agent stops and flags it. Default is 3. Set it lower (1 or 2) for simple skills. Set it higher (5) for complex ones.