The StackThe TrapRight ApproachFind the ProblemYour PathThe JourneyFramework Reference
Builders Guide

Building Products
with AI Agents

You've got the mental model — the cognitive loop, MCP, Skills, memory, guardrails. Now the question is: what do you build with? 30+ tools across 7 layers, from brainstorming to enterprise deployment. But picking a tool first is the most expensive mistake you can make.

← Back to the Primer Explore the Toolkit →
01 — The Excitement

From Passion Project to Enterprise Product

The stack covers every stage. Each layer is a distinct type of thing. Know what each one IS — and where it fits in your journey.

THE BUILDER'S STACK IDEA → PROTOTYPE → PRODUCT → SCALE AI ASSISTANTS Chat interfaces you think, brainstorm & research with Claude.ai ChatGPT Gemini Perplexity AI APP BUILDERS Prompt-to-deployed-app — no coding required Lovable Bolt.new Base44 Replit v0 (Vercel) AI CODE EDITORS & AGENTS IDEs with AI built in + autonomous coding collaborators Cursor Windsurf Claude Code Copilot Devin AGENT FRAMEWORKS & SDKs Libraries & SDKs for building autonomous AI agents LangGraph CrewAI AutoGen Semantic K. Google ADK Strands SDK PROTOCOLS & STANDARDS Open specifications — the connective tissue of the agent economy MCP A2A (Google) Skills / Plugins ENTERPRISE CLOUD PLATFORMS Managed, governed, production-scale agent services Azure Foundry AWS Bedrock Vertex AI NemoClaw AUTOMATION & INTEGRATION Connect everything — workflow platforms & iPaaS Workato Power Automate Zapier / Make n8n
AI Assistants — Chat Interfaces You Think With
CL

Claude.ai

Anthropic's frontier AI. Extended thinking, artifacts for instant apps, project knowledge, web search. The thinking partner behind this page.

Opus 4.6200K context
GP

ChatGPT

OpenAI's conversational AI. GPT-4o multimodal, Canvas, DALL-E, plugins. 200M+ weekly users.

200M+ usersGPT-4o
GM

Gemini

Google's AI. 2M token context, Workspace integration, multimodal. Strongest in Gmail, Docs, Sheets.

2M tokensWorkspace
PX

Perplexity

AI research engine. Cited, sourced answers. Deep Research writes full reports. Replaces "let me Google that."

CitedDeep Research
AI App Builders — Prompt to Deployed Product

Describe what you want. Get a working app in minutes. This is where passion projects become real.

LV

Lovable

Market leader. React + TypeScript, Supabase backend, GitHub sync, 100+ integrations. Design-first.

$300M ARRReact/TS
BN

Bolt.new

Full-stack with real in-browser IDE. File tree, terminal, NPM. Open-source engine. Fastest for prototyping.

Full IDEOpen-source
B4

Base44

Zero-setup backend. Auto-generates DB, auth, RBAC, hosting. Multi-model. Acquired by Wix for $80M.

$80M WixZero-setup
RP

Replit

Cloud IDE + Agent 3 autonomous build/test/deploy. Design Mode. Connectors for 20+ services.

$9B valAgent 3
V0

v0 by Vercel

Text-to-UI for React + Tailwind. Cleanest frontend code. Components for Next.js projects.

ReactTailwind
AI Code Editors & Agents — Prototype to Production

When prototypes need real engineering. AI-native IDEs and autonomous coding agents.

CU

Cursor

VS Code fork. Composer for multi-file generation, 8 parallel agents, Plan Mode, visual editor.

$2B+ ARR8 agents
WS

Windsurf

#1 AI IDE. Cascade flow, Arena Mode. Merging with Devin for fully AI-driven development.

#1 rankedCascade
CC

Claude Code

Terminal-native agent. Reads codebase, runs commands, manages Git, spawns sub-agents. 1M context.

4% commits1M ctx
GC

GitHub Copilot

Most adopted. Inline completions, chat, agent mode. VS Code, JetBrains, Vim. 26M+ users, $10/mo.

26M+$10/mo
DV

Devin

Autonomous software engineer. Plans, writes, tests, deploys via Slack. Sandboxed. $4B valuation.

$4BAutonomous
Agent Frameworks — Build Autonomous Systems

Libraries for AI agents that reason, use tools, execute multi-step tasks. Open-source orchestration + cloud provider SDKs.

LG

LangGraph

Graph-based state machine. Nodes, edges, loops. Durable execution. Human-in-the-loop. Strong Planning + Observe loop from the cognitive architecture.

44.6K stars12M+ dl
CA

CrewAI

Role-based multi-agent. Agents with roles, goals, backstories collaborate as a team. The A2A delegation pattern in action. 40% faster to prod.

45.9K stars100K+ devs
AG

AutoGen

Microsoft's multi-agent framework. Conversational agent patterns, code execution, group chat orchestration. Strong Reasoning layer.

40K+ starsMicrosoft
SK

Semantic Kernel

Microsoft's SDK for AI integration. C#/Python/Java. Deep Azure + M365 + Copilot integration. Enterprise-first.

C#/Py/JavaAzure native
GK

Google ADK

Agent Development Kit. Build agents with Gemini, deploy to Vertex AI Agent Engine. Native A2A support.

GeminiVertex AI
SS

Strands SDK

AWS's agent SDK. Model-driven, built for Bedrock. MCP-native tool connectivity. Open-source.

AWSMCP native
OC

OpenClaw

Open-source agent runtime. Runs locally, connects via chat apps. 250K+ stars. All 8 building blocks running on your machine — MCP native, 100+ Skills, dual-layer memory.

250K+ stars50+ integrations
LS

LangSmith

Observability for agents. Trace every decision, tool call, token. Debug why agents do what they do.

5K free/mo$39/seat
Protocols & Standards — The Connective Tissue

Not products — open specifications. The reason all tools above work together.

MC

MCP

Model Context Protocol. Connects AI to tools, databases, APIs. Anthropic, OpenAI, Google, Microsoft. The USB of AI.

6,400+ servers97M SDK/mo
A2

A2A Protocol

Google's Agent-to-Agent. MCP = agents to tools. A2A = agents to agents. Communication backbone.

GoogleMCP complement
SK

Skills & Plugins

Modular capabilities. SKILL.md files with instructions and tool definitions. Build once, share everywhere.

100+Open
Enterprise Cloud — Managed, Governed, at Scale

SSO, compliance, audit trails, VPC isolation. Managed services from cloud providers.

MF

Microsoft Foundry

Azure unified AI. 11K+ models, 1,400+ connectors, MCP, Entra ID. Agent Service GA. Deploy to M365 Copilot.

11K+ models1,400+
AB

AWS Bedrock

Fully managed agents. Select model, write instructions. IAM/VPC/CloudWatch native. AgentCore for scale.

AWS nativeAgentCore
VA

Vertex AI

Google's agent platform. ADK in <100 lines Python. A2A interop. Agent Engine managed deploy.

ADKA2A native
NC

NemoClaw

Enterprise OpenClaw + Nvidia guardrails. Nemotron models, DGX Spark/Station. Always-on local agents.

NeMoDGX
Automation & Integration — Connect Everything

Workflow platforms connecting agents to SAP, Salesforce, Slack, Dynamics, 1,000+ more.

WK

Workato

Enterprise iPaaS. 1,000+ connectors, recipe-based automation, governance. Bridge to enterprise systems.

1,000+Enterprise
PA

Power Automate

Microsoft low-code. Visual flows, Copilot integration, SharePoint/Teams/Dynamics. Best for M365 orgs.

M365Low-code
ZP

Zapier / Make

Lightweight SaaS automation. No-code, fast, cheap. Simple triggers and app-to-app glue.

No-codeQuick
N8

n8n

Open-source workflow automation. Self-hosted, full data control, visual node editor. AI nodes built in.

Open-sourceSelf-hosted
02 — The Trap

Seven Layers of Seduction

That stack you just saw? It's intoxicating. Thirty-plus tools, each more impressive than the last. You can go from a sentence to a deployed app in four minutes. You can run a personal AI agent from your laptop that manages your email while you sleep. The temptation is overwhelming: grab a tool and start building.

Don't.

The pattern: Someone discovers OpenClaw or Claude Code or Lovable, asks "what can I build?" Weeks of technically impressive work nobody wants. Applause at the demo. Zero users after.

Tool-first = demos, not products. Solutions looking for problems. Features nobody asked for. Architectures chosen because they're interesting. The nineteen out of twenty agent pilots that never reach production? Most of them started here.

Building the wrong thing is now faster too. Speed amplifies good decisions and bad ones equally. A tool that lets you ship in 4 minutes also lets you waste 4 months on the wrong problem in 4 minutes.

The discipline: Start with a problem — a real person with a real pain point they'd pay to solve. Validate the use case. Calculate the ROI. Then — only then — pick the tool from the stack. That sequence is the difference between the 1 in 20 that makes it to production and the 19 that don't.

Tool-First Death Spiral

  • "This tech is amazing, let me find a use"
  • Build something technically impressive
  • Demo gets applause from builders
  • Launch to crickets
  • Pivot repeatedly
  • Conclude "market wasn't ready"

What Actually Works

  • "This person has a painful problem"
  • Validate they'd pay to solve it
  • Calculate the ROI
  • Select the tool that fits
  • Build the minimum
  • Iterate on real feedback
03 — The Right Approach

Problem → Use Case → ROI → Tool

Flip the sequence. Start with a real problem. Validate the use case. Calculate ROI. Then pick the tool. Every step earns the right to take the next one.

01

Find a Real Problem

Talk to 20 people. Actual pain — wasted time, lost money, missed opportunities. Not what you think is cool — what they'd pay to fix.

02

Define the Use Case

One sentence: "[Person] wastes [cost] doing [workaround]." If you can't say it in one sentence, you don't understand it yet.

03

Prioritize with Data

Impact × feasibility × risk. Not gut feel — structured evaluation. The Use Case Discovery Workbook scores across 10 dimensions with risk-adjusted recommendations.

04

Calculate the ROI

Time saved + errors avoided + revenue unlocked = value. API + infra + your time = cost. Aim for 5x+ ROI. If below 3x, you need a bigger problem, not a better tool.

05

Pick the Tool

Now — only now — match requirements to the right path and layer from the stack. Your maturity level determines your path.

06

Build the Minimum

Prototype in days, not months. Show to those 20 people. Iterate on reality, not assumptions.

Live Exercise

Think of the last time you saw someone doing manual, repetitive work:

  1. What were they doing, step by step?
  2. How many hours per week?
  3. What's the hourly cost?
  4. What breaks when done manually?

You just found a use case — and the start of an ROI calculation.

04 — Finding the Problem

Where AI Agents Create Real Value

Not every problem needs AI. Focus where agents have a genuine advantage. The Use Case Discovery & Prioritization Workbook in the Toolkit systematizes this with 10 agent-specific patterns and a risk-adjusted scoring framework.

Manual processes. Copying data between systems, following checklists, pattern recognition on unstructured data — agent opportunities. Think of the O2C cycle from the Primer: 4 cognitive loops across order entry, fulfillment, invoicing, and cash collection.

Domain expertise bottlenecks. One expert reviewing contracts or triaging tickets — encode that knowledge into Skills so the agent can handle the 80% and escalate the 20%.

Integration nightmares. Data in silos, disconnected tools. MCP was built for this — one protocol connecting your agent to every system.

Vertical opportunities. Specific industry + core workflow rebuilt with agents = products incumbents can't replicate. Healthcare, SLED, financial services, SaaS — see the Agent Showcase for 11 patterns.

Problem Litmus Test

  • Describe in one sentence?
  • Would someone pay today?
  • Current solution painful enough to switch?
  • Can you reach 100 people?
  • AI gives 10x advantage over manual?
  • Recurring, not one-time?
  • Which maturity level? (Start Level 1)
05 — Your Problem Determines Your Path

Five Paths Through the Stack

You've identified the problem and validated the ROI. Now — which layers of the stack do you actually need? Nobody uses all 7. Mapped to the maturity levels from the Primer.

PATH 01

Quick MVP · Level 1

"I want it live this weekend." Claude.ai → Lovable/Bolt → Zapier (optional). 2–3 layers only. Information agents, simple prototypes.

PATH 02

Workflow Automation · Level 2

"Automate invoices, onboarding." Claude → Cursor/CC → MCP → Workato/n8n. Action agents with read + write. Add governance.

PATH 03

AI Agent Product · Level 2–3

"Build a testing/evaluation agent." Claude → Claude Code → CrewAI/LangGraph → MCP → Cloud deploy. Full cognitive loop with Skills.

PATH 04

Personal Agent · Level 1–2

"Manage my calendar, email, tasks." Claude → OpenClaw → Skills/MCP. All 8 building blocks running locally. Memory + cron.

PATH 05

Enterprise Multi-Agent · Level 3

"Clinical trials, compliance, O2C." Claude → Cursor/CC → LangGraph → MCP+A2A → Foundry/Bedrock → Workato. All 7 layers. Full governance required.

You've Found the Problem. You Know Your Path. Now Execute With Discipline.

This is where most teams jump straight to the IDE. The ones that succeed don't. They start with the Agentic Engineering lifecycle — a 6-stage operational framework: Justify → Architect → Govern → Build → Gate → Operate. And the first tool in that lifecycle is the Use Case Discovery & Prioritization Workbook — 10 agent-specific evaluation dimensions, risk-adjusted scoring, and a Selection Brief that tells you whether the math supports building.

Explore the Toolkit → Start with the Workbook →
06 — The Path Forward

The Builder's Journey

The Primer gave you the mental model. This page gave you the tool selection discipline. The Toolkit gives you the operational playbooks. Here's the full sequence.

Talk to 20 People

Real humans, real problems, real budgets. Not what you think is cool — what they'd pay to fix.

Define Use Case + ROI

One sentence. Napkin math. Use the Prioritization Workbook to score across 10 dimensions. If ROI isn't 5x+, find a bigger problem.

Pick Your Path

Match problem to maturity level. Choose from the 5 paths. You don't need all 7 layers.

Govern Before You Build

Governance Policy Template (32 pre-production gates) and Identity & Trust Template (49 security controls) — before the first line of code.

Build & Integrate

Now open the IDE. Select your framework from the reference table below. Connect MCP servers. Encode expertise into Skills.

Gate & Validate

Pre-production review. Does the agent do what it should? Does it NOT do what it shouldn't? The observation step from the cognitive loop — applied to your release process.

Operate & Improve

Incident response, monitoring, continuous improvement. The same Sense → Think → Plan → Act → Observe loop — applied to your operations.

The Three Pages, One Story

The Primer gave you the mental model — the noise filter for evaluating every agent product and framework. This page gave you the discipline — problem first, tool second. The Toolkit gives you the operational playbooks to execute — 28 tools organized across the 6-stage Agentic Engineering lifecycle. 9 available for download today. Start with the Use Case Discovery & Prioritization Workbook.

Problem. Use Case. ROI. Path. Governance. Then Build.

That's the sequence. The Toolkit has the playbooks for every step.

Browse the Toolkit → See Agent Architectures →
Reference

The Framework Comparison

You've chosen your path and started the Toolkit lifecycle. When you reach the Build & Integrate stage, this is your framework selection reference — the tools mapped to use cases, with the building blocks each one emphasizes.

Problem identified, ROI validated. Map requirements to the stack.

If you need...ConsiderWhyBest for
BrainstormingClaude.ai / ChatGPT / GeminiReasoning, search, artifactsIdeas, planning
First prototypeLovable / Bolt.new / Base44Prompt → app in minutesMVPs, validation
UI componentsv0 by VercelReact + Tailwind from textDashboards, pages
Scaling past MVPCursor + Claude CodeAI IDE + terminal agentProduction code
Tool connectivityMCP ServersOne protocol, all modelsData & API integration
Personal agentsOpenClawLocal, extensible, chat-nativeProductivity
Complex workflowsLangGraphState graphs, durable execution, loopsFinance, healthcare, logistics
Fast multi-agentCrewAIRole-based, 40% faster to prodContent, triage, research
Conversational agentsAutoGenMulti-agent chat, code executionCollaborative reasoning
Microsoft ecosystemSemantic KernelC#/Python/Java, M365 + Copilot native.NET enterprise shops
Google ecosystemGoogle ADKGemini-native, A2A built-inGCP-first organizations
AWS ecosystemStrands SDKBedrock-native, MCP built-inAWS-first organizations
Enterprise AzureAzure AI Foundry1,400 connectors, Entra IDRegulated industries
Enterprise AWSBedrock + AgentCoreIAM, VPC, CloudWatchAWS organizations
Enterprise GCPVertex AI + Agent EngineADK, A2A, Agent GardenGoogle Cloud orgs
Agent observabilityLangSmith / LangFuseTrace, debug, evaluate agentsAny framework
Enterprise iPaaSWorkato / Power Automate1,000+ connectorsSAP, Salesforce
Simple automationsZapier / Make / n8nNo-code, fast, cheapTriggers, flows
Launchpad 🧠 Primer 🔧 Builder’s Guide 🧰 Toolkit
From the Book

This topic is covered in depth in The Agentic Enterprise Strategy — the complete practitioner’s guide to architecting, governing, and operating AI agent systems in production.

Get the Book Browse Toolkit