AI Agent Builder#

Structured framework for building AI agents that work in production. Based on the Storm & Storm methodology.

When to Use#

• Designing a new AI agent from scratch
• Planning architecture for an automated workflow
• Reviewing or improving an existing agent's design
• Teaching someone how to build agents

The 8-Step Process#

Follow these steps in order. Each step has a clear goal and concrete deliverables.

Step 1: Choose a Task#

Pick ONE painful, repeating workflow. Not "AI in general."

• Must be repeatable (weekly+), follow steps, have clear I/O
• Define success: "Given X, the agent should output Y so that Z happens."

Step 2: Map the Steps#

Break the task into 4–7 steps: INPUT → ACTIONS → DECISION → OUTPUT

• Classify each step: ⚖️ pure rules | 📖 heavy reading/writing | 🎯 judgement calls
• Choose infrastructure (no-code vs dev-friendly)
• You need: strong model + tool calling + basic logs

Step 3: Specify Inputs, Outputs & Tools#

Treat the agent like an API, not a chatbot.

• Define required input fields (text, file, URL, ID)
• Define structured outputs (JSON/template the system can trust)
• Attach tools: data (search/DB/CRM), action (email/Slack/tasks), orchestration (schedulers/webhooks/queues)

Step 4: Write the System Prompt#

Create a clear role with: role definition, boundaries, style, 1–2 example conversations.

• Use ReAct pattern: observe → think → act → reflect

Step 5: Add Memory#

Three layers: conversation state, task memory, knowledge memory (vector store/file search).

• Key question: "What does this agent need to remember for the next step to be smarter?"

Step 6: Add Safeguards#

Gate high-risk actions (email, data changes, money) behind human approval.

• Rules: never invent IDs, ask when ambiguous
• Log every tool call and decision for audit

Step 7: Build the Interface#

Match to where users work: chat, Slack command, button in app, or web form.

Step 8: Test#

For each real example: watch the trace, score correctness + efficiency + time saved.

• Tighten prompts/tools/rules where it fails. Iterate.

Detailed Reference#

For expanded details on each step, including selection criteria, classification examples, tool categories, memory layer patterns, and a pre-launch checklist:

→ Read references/guide.md

Output Format#

When using this framework to design an agent, produce a design document covering:

text

# Agent Design: [Name]

## Task & Success Criteria
[Step 1 output]

## Step Map
[Step 2 output — numbered steps with classifications]

## I/O Specification
[Step 3 output — inputs, outputs, tools]

## System Prompt
[Step 4 output — the actual prompt]

## Memory Architecture
[Step 5 output — which layers, what's stored]

## Safeguards
[Step 6 output — gated actions, rules, logging]

## Interface
[Step 7 output — chosen interface and why]

## Test Plan
[Step 8 output — example inputs, expected outputs, scoring criteria]