# Matthew Kruczek

> EY Managing Director writing practitioner-depth articles on AI agent architecture, MCP protocols, and scaling enterprise AI from pilot to production.

This is the llms.txt index for https://matthewkruczek.ai. The site is a personal thought-leadership blog covering enterprise AI strategy, agentic architecture, the Model Context Protocol, and AI readiness for large organisations.

## Key pages

- [Home](https://matthewkruczek.ai/): Profile, expertise, speaking topics, and projects.
- [All articles](https://matthewkruczek.ai/blog.html): Full blog index with category filters.
- [RSS feed](https://matthewkruczek.ai/feed.xml): Machine-readable feed of all articles.
- [Privacy policy](https://matthewkruczek.ai/privacy.html): How data is handled.

## Articles

- [Copilot Is Not a Product: A CTO's Map | Matthew Kruczek](https://matthewkruczek.ai/blog/copilot-is-not-a-product.html): Microsoft put one name on a dozen products with different buyers, bills and prerequisites. The five layers, what each inherits, and what you already pay for.
- [The Memory: You Only Built Half of It | Matthew Kruczek](https://matthewkruczek.ai/blog/agentic-memory-and-state.html): Memory is what your agent knows. State is where it is. Lose one and it forgets a fact, lose the other and it pays a departing employee twice. Anatomy Part 4.
- [The Bloodstream: Your Tools Are the Risk | Matthew Kruczek](https://matthewkruczek.ai/blog/agentic-gateway-tools.html): The model is a commodity behind a gateway. The tools reach into systems you do not control, and neither one shows up in a demo. Part 3 of the Anatomy series.
- [The Brain Stem: Four Multi-Agent Shapes | Matthew Kruczek](https://matthewkruczek.ai/blog/agentic-orchestration-runtime.html): Every team says they built a multi-agent system. Few can name which of the four they run, because the demo chose for them. Part 2 of the Anatomy series.
- [The Front Door: Your Agent Has No Identity | Matthew Kruczek](https://matthewkruczek.ai/blog/agentic-front-door-identity.html): Your agent acts on a borrowed token, so the audit log names a human who was asleep. Agent identity is a layer, not a checkbox, and here is how you build it.
- [The Anatomy of an Agentic System | Matthew Kruczek](https://matthewkruczek.ai/blog/anatomy-of-an-agentic-system.html): Gartner says 40% of agentic AI projects die by 2027, and the model is never the cause. A production agentic system is eleven layers, not four. Here is the map.
- [What Is an AI Second Brain? A Plain Guide | Matthew Kruczek](https://matthewkruczek.ai/blog/what-is-an-ai-second-brain.html): A plain-English walkthrough of the AI second brain: what it is, how the machine works, the problem it solves, and how to tell if you or your company needs one.
- [Tokenomics in Practice: Govern AI Spend | Matthew Kruczek](https://matthewkruczek.ai/blog/tokenomics-in-practice.html): The exact 90-day system for governing enterprise AI token spend: one gateway, cost per outcome, funded efficiency, and soft budgets instead of hard caps.
- [Why MD Files Will Never Go Out of Style | Matthew Kruczek](https://matthewkruczek.ai/blog/md-files-never-go-out-of-style.html): Your agent stack changes at three speeds. Models churn in weeks, harnesses in months, your markdown intent files barely at all. Invest where work lasts.
- [What Is a Forward Deployed Engineer? | Matthew Kruczek](https://matthewkruczek.ai/blog/what-is-an-fde-anyway.html): Everyone hires Forward Deployed Engineers and every firm sells them. A definition that means something: value that compounds into the firm and the client.
- [Yes, You Can Run Claude on Azure | Matthew Kruczek](https://matthewkruczek.ai/blog/claude-on-azure.html): You can deploy Anthropic's Claude inside Microsoft Foundry today, then govern it with the same Azure API Management gateway that already controls your OpenAI traffic.
- [Just Build It in AI | Matthew Kruczek](https://matthewkruczek.ai/blog/just-build-it-in-ai.html): The code got cheap. Delivery didn't. Why "just build it in AI" prices the one part of software that was already free, and where the real cost actually moved.
- [You're Not Writing Prompts Anymore. You're Writing Loops. | Matthew Kruczek](https://matthewkruczek.ai/blog/writing-loops-not-prompts.html): Boris Cherny says he writes loops, not prompts. The shift from one-off instructions to self-running agents, explained in plain English, and the four things that become your job.
- [Your Agent Isn't Learning. It's Taking Notes. | Matthew Kruczek](https://matthewkruczek.ai/blog/agent-isnt-learning-taking-notes.html): Every vendor sells a "self-learning" agent. Almost none have one. Here is the mechanism behind real self-improvement, and the five-second test that exposes the difference.
- [The Trust Boundary Moved: APIM as Your Agentic Governance Layer | Matthew Kruczek](https://matthewkruczek.ai/blog/apim-agentic-governance.html): Model-level safety can't govern agent behavior. The API gateway is where model calls, tool invocations, and agent-to-agent traffic converge into one chokepoint you can govern.
- [You Installed Claude. Now the Hard Work Starts | Matthew Kruczek](https://matthewkruczek.ai/blog/enterprise-ai-governance.html): The enterprise AI license is the easy part. A practical playbook for the governance that follows: shadow AI, data classification, access controls, and token spend.
- [Tokenomics: Why Capping AI Token Spend Backfires | Matthew Kruczek](https://matthewkruczek.ai/blog/tokenomics-beyond-the-cap.html): A hard cap on AI token spend optimizes the wrong number. A practical framework for governing token economics by cost per outcome, without cutting developers off.
- [If I Built My Engineering Organization From Scratch Today | Matthew Kruczek](https://matthewkruczek.ai/blog/agentic-engineering-org.html): Your developers got faster. Your company did not. The hard part moved from writing code to checking it, and most org charts never moved with it.
- [The Logistics of Agentic Development: The Part Nobody Plans For | Matthew Kruczek](https://matthewkruczek.ai/blog/agentic-development-logistics.html): Everyone writes the case for agentic development. Almost no one writes the operating manual. Five artifacts that turn a decision to go agentic into a team that actually runs.
- [Why Your Agent Forgets the Rules: The Neurosymbolic Case for the Harness | Matthew Kruczek](https://matthewkruczek.ai/blog/neurosymbolic-harness.html): You wrote the rule down, added ALWAYS and CRITICAL, and the agent ignored it anyway. That is not a prompting failure. Neurosymbolic AI explains why.
- [From Pilot to Production: The Scaling Playbook | Matthew Kruczek](https://matthewkruczek.ai/blog/scaling-pilot-to-production.html): 88% of organizations use AI. Fewer than 10% have scaled it. The 90-day sprint, five failure mode diagnostics, and industry playbooks for bridging the pilot-to-production gap.
- [The RAG Taxonomy: Why One Retrieval Pipeline Can't Serve Every Enterprise Query | Matthew Kruczek](https://matthewkruczek.ai/blog/rag-taxonomy-enterprise.html): RAG is not one technology. It is five distinct architectures. Most enterprises built the 2023 version and never upgraded. Here is the taxonomy that fixes it.
- [Harness Engineering Playbook: Match AI Methodology to Project Rigor | Matthew Kruczek](https://matthewkruczek.ai/blog/building-agentic-harness.html): 15 SDD frameworks span 3 rigor levels. A harness engineering guide and Claude Code plugin that routes your team to the right AI coding methodology.
- [Engineering AI Enablement: Beyond Tool Licenses | Matthew Kruczek](https://matthewkruczek.ai/blog/engineering-ai-enablement.html): 74% of AI tool buyers can't show business value. The gap isn't tools, it's the platform layer: shared infrastructure, golden paths, observability, cost governance, and developer experience.
- [Governed AI at Speed: Why Most AI Oversight Kills Innovation Before It Starts | Matthew Kruczek](https://matthewkruczek.ai/blog/governed-ai-at-speed.html): Three governance models, a three-tier risk classification, and a 90-day sprint for building AI oversight that protects without paralyzing your AI program.
- [Agentic Memory: Why Bigger Context Windows Won't Save Your Enterprise AI | Matthew Kruczek](https://matthewkruczek.ai/blog/agentic-memory-enterprise-ai.html): Your AI agents forget everything when the session ends. Here's the architecture that fixes it, and why throwing more context at the problem makes it worse.
- [AI Literacy at Scale: Why Training Programs Fail and What Works Instead | Matthew Kruczek](https://matthewkruczek.ai/blog/ai-literacy-skills-strategy.html): Four AI capability levels every role maps to, the failure modes that turn training into waste, and a 90-day skills sprint that builds measurable capability.
- [The Token Tax: Why Untrained Developers Are Your Most Expensive AI Problem | Matthew Kruczek](https://matthewkruczek.ai/blog/token-tax-developer-training.html): Untrained developers waste hundreds of thousands in AI tokens. How token literacy, open source monitoring, and architectural discipline change the equation.
- [The AI Operating Model: Why Organizational Structure Determines Everything Downstream | Matthew Kruczek](https://matthewkruczek.ai/blog/ai-operating-model.html): Three AI operating models determine whether your organization can scale past pilot. A diagnostic for identifying which you're running and when to transition.
- [Your Data Is Not AI-Ready: The Six-Month Workstream That Fixes It | Matthew Kruczek](https://matthewkruczek.ai/blog/data-readiness-workstream.html): Data readiness kills 60% of AI projects. A six-month parallel workstream to fix data access, quality, and governance before your AI program stalls.
- [Your AI Roadmap Won't Survive the Next Budget Review | Matthew Kruczek](https://matthewkruczek.ai/blog/ai-strategic-alignment.html): Most AI roadmaps are technology wish lists that collapse at the first budget review. A prioritization framework and 90-day test for building strategic AI alignment that holds.
- [The AI Executive Sponsor: Who They Should Be and Why Most Companies Get It Wrong | Matthew Kruczek](https://matthewkruczek.ai/blog/ai-executive-sponsor.html): More than half of AI initiatives lose their executive sponsor within six months. Five anti-patterns reveal why, and a 30-day diagnostic tells you if yours is working.
- [The AI Readiness Scorecard: Eight Dimensions That Determine Whether You Scale or Stall | Matthew Kruczek](https://matthewkruczek.ai/blog/ai-readiness-scorecard.html): Only 13% of organizations are ready to deploy AI at scale. An eight-dimension scorecard reveals where yours is breaking down and which gaps to close first.
- [Same Prompt, Different Results. Your Agent Harness Is the Multiplier. | Matthew Kruczek](https://matthewkruczek.ai/blog/harness-is-the-multiplier.html): Most engineers leave 60-70% of their AI agent's capability on the table. A production harness with rules, hooks, skills, and memory changes everything.
- [Microsoft Just Made AI Agents an Engineering Discipline | Matthew Kruczek](https://matthewkruczek.ai/blog/microsoft-agent-engineering-discipline.html): Microsoft's 2026 Frontier Stack treats AI agents as engineering artifacts, not demos. Build, Measure, Govern: the reference architecture enterprise teams have been waiting for.
- [The Missing Layer: Why Your AI Agents Need a Package Manager | Matthew Kruczek](https://matthewkruczek.ai/blog/the-missing-layer.html): Microsoft's open-source APM applies the package manager pattern to AI agent configuration. One manifest, lock files, reproducible setups across your org.
- [AI Vocabulary Guide for Executives: 24 Terms You Need | Matthew Kruczek](https://matthewkruczek.ai/blog/ai-vocabulary-executives.html): Your engineering team is speaking a new language. A plain-English guide to 24 essential AI terms, organized for business leaders making investment decisions.
- [Measuring Agentic AI Impact: The Three-Layer Framework | Matthew Kruczek](https://matthewkruczek.ai/blog/measuring-agentic-ai-impact.html): Most enterprises measure agentic AI with RPA-era KPIs. Here's the three-layer framework leaders actually need: operational, decisional, and adaptive capacity metrics.
- [Agent Harnesses Need Fewer Layers, Not More | Matthew Kruczek](https://matthewkruczek.ai/blog/agent-harnesses-fewer-layers.html): Vercel cut 15 tools to 2 and hit 100% accuracy. Evidence from Vercel, Manus, and OpenAI shows simpler agent harnesses consistently outperform complex ones.
- [CLI vs. MCP: Here Is How to Think About It | Matthew Kruczek](https://matthewkruczek.ai/blog/cli-vs-mcp-debate.html): CLI and MCP are not competing technologies. They solve different problems at different layers of your AI architecture. Here's how to think about which one belongs where.
- [I Built the Progressive Disclosure Hub. Here's What the Benchmarks Say. | Matthew Kruczek](https://matthewkruczek.ai/blog/fastmcp-codemode-hub-benchmarked.html): FastMCP 3.1's CodeMode collapses 28 tools into 3, cutting schema tokens 91% on every session start. Real benchmarks, real trade-offs, and the API details that changed.
- [Stop Learning Agent Frameworks. Learn Context Windows First. | Matthew Kruczek](https://matthewkruczek.ai/blog/context-windows-before-agent-frameworks.html): Every AI agent failure traces back to context window mismanagement. Master context windows before you touch Microsoft Agent Framework or any other orchestration layer.
- [Enterprise AI Needs 3 Layers: MCP, Skills & A2A Explained](https://matthewkruczek.ai/blog/mcp-skills-a2a-three-layers.html): Most enterprise AI stacks are missing a layer. MCP = connectivity, Skills = knowledge, A2A = collaboration. Full architecture guide with implementation.
- [Computer-Using Agents in Azure AI Foundry: A Hands-On Guide | Matthew Kruczek](https://matthewkruczek.ai/blog/computer-using-agents-foundry.html): Computer-Using Agents represent a fundamental shift in enterprise automation. Here's what CTOs need to know about CUA in Azure AI Foundry vs. traditional RPA.
- [C# vs Python for Enterprise AI: The Honest Comparison | Matthew Kruczek](https://matthewkruczek.ai/blog/csharp-python-enterprise-ai.html): 6-8 million .NET developers are discovering C# may be the better choice for production AI systems. Here's the honest comparison for enterprise teams choosing their AI stack.
- [Progressive Disclosure MCP: 85x Token Savings Benchmark](https://matthewkruczek.ai/blog/progressive-disclosure-mcp-servers.html): I benchmarked progressive disclosure vs flat MCP loading across 5 scenarios. Result: 85-100x token reduction with higher agent accuracy. Data and implementation guide inside.
- [How to Build a .NET Skills Executor for Claude AI Agents | Matthew Kruczek](https://matthewkruczek.ai/blog/dotnet-skills-executor.html): Build an AI Skills Executor in .NET that packages your organization's procedural knowledge into composable, portable skills using Azure AI Foundry and the MCP C# SDK.
- [4 Multi-Agent Patterns Microsoft Uses in Production | Matthew Kruczek](https://matthewkruczek.ai/blog/multi-agent-patterns-microsoft.html): Translating Google's ADK multi-agent patterns into production-ready Python code using Microsoft Agent Framework. Learn sequential, parallel, critic, and routing patterns.
- [Agent Skills: What AI Agents Need Beyond RAG and Tools](https://matthewkruczek.ai/blog/agent-skills-missing-link.html): RAG and tool-use got agents reading and acting, but procedural knowledge is the gap. How Anthropic's Skills framework gives agents the how, not just the what.
- [How to Stay Relevant as a Developer in the AI Era | Matthew Kruczek](https://matthewkruczek.ai/blog/programming-success-ai-era.html): 10 pieces of career advice for software developers from a 25-year industry veteran. The real-world skills that separate developers who thrive from those who plateau.
- [How We Cut AI Token Costs 85% with One Design Pattern | Matthew Kruczek](https://matthewkruczek.ai/blog/toon-token-optimization.html): Token-Oriented Object Notation (TOON) reduces AI token usage by 40-50% while improving accuracy. A practical guide to enterprise AI optimization with real-world benchmarks.
- [Getting Real Business Value from Microsoft Copilot | Matthew Kruczek](https://matthewkruczek.ai/blog/copilot-enterprise-value.html): A strategic three-stage roadmap for Microsoft Copilot adoption—from quick wins to full AI transformation. Learn how to maximize ROI across your enterprise.
- [How AI Agents Are Transforming App Modernization Projects | Matthew Kruczek](https://matthewkruczek.ai/blog/app-modernization-sdlc.html): Your modernization strategy is already obsolete. Learn how to integrate application modernization with agent-powered development for 10x competitive advantage.
- [AI Agents in Your SDLC: A Crawl-Walk-Run Playbook](https://matthewkruczek.ai/blog/agent-first-getting-started.html): Step-by-step framework for embedding AI agents into software delivery, from single-developer copilots to autonomous pipelines—with real cycle-time benchmarks.
- [MCP-UI: Breaking the Text Barrier in Enterprise AI | Matthew Kruczek](https://matthewkruczek.ai/blog/mcp-ui-enterprise.html): Why interactive interfaces are the next competitive advantage. MCP-UI extends the Model Context Protocol to deliver visual, actionable AI experiences at enterprise scale.
- [Training Engineers to Orchestrate AI Agents, Not Just Prompt | Matthew Kruczek](https://matthewkruczek.ai/blog/training-engineers-orchestrate.html): The role of software developer is shifting from code implementation to system orchestration. Learn how to train developers for the AI-first future—at every level.
- [Why Clear Requirements Give You an Edge in the AI Era | Matthew Kruczek](https://matthewkruczek.ai/blog/clear-requirements-ai-advantage.html): The shift to AI-driven operations isn't about better algorithms—it's about mastering specification. Organizations that excel at translating intent into precision will dominate.
- [Starting Your AI Strategy From Scratch? Read This First | Matthew Kruczek](https://matthewkruczek.ai/blog/starting-fresh-ai-strategy.html): Companies that will dominate the next decade won't incrementally adopt AI—they'll rebuild operations from the ground up around agent-first principles. Here's how.
- [Context Engineering: The Skill That Separates Good AI From Great AI | Matthew Kruczek](https://matthewkruczek.ai/blog/context-engineering-enterprise-ai.html): Context engineering is the critical differentiator between AI that delivers transformative value and AI that disappoints. Here's how to master it with Microsoft's AI platform.
- [Agent-First Enterprise: Architecting Beyond Copilots](https://matthewkruczek.ai/blog/agent-first-enterprise.html): Copilots help individuals. Agent-first architecture restructures workflows around autonomous AI—a practical reference model for teams ready to scale past pilots.
- [AI Adoption Paradox: Why 95% of AI Pilots Fail](https://matthewkruczek.ai/blog/ai-adoption-paradox.html): Most enterprises adopt AI everywhere yet ship nothing to production. Seven principles explain what the successful 6% do differently—and how to join them.
- [AI Bots That Actually Reach Customers (Not Just FAQs) | Matthew Kruczek](https://matthewkruczek.ai/blog/bots-reaching-customers.html): Most enterprise chatbots fail because they're FAQ machines. Learn how AI-native bots proactively reach customers across channels and drive real business outcomes.
- [How AI Agents Are Rewriting Software Development in 2026 | Matthew Kruczek](https://matthewkruczek.ai/blog/agents-changing-software-development.html): AI agents are fundamentally changing how software gets built—from code generation to testing to deployment. Here's what every engineering leader needs to know.
- [How AI Is Rewriting Every Phase of Software Development | Matthew Kruczek](https://matthewkruczek.ai/blog/transforming-sdlc-with-ai.html): AI isn't just speeding up software development—it's fundamentally changing every phase of the SDLC. Here's what engineering leaders need to know and do right now.
- [How to Choose an AI Framework Without Getting It Wrong | Matthew Kruczek](https://matthewkruczek.ai/blog/choosing-ai-frameworks.html): Picking the wrong AI framework wastes months and millions. Here's the decision framework enterprise leaders use to evaluate Semantic Kernel, LangChain, Agent Framework, and more.
- [How to Select AI Tools Without Wasting 6 Months and $500K | Matthew Kruczek](https://matthewkruczek.ai/blog/selecting-ai-tools.html): Most organizations waste 6-18 months evaluating AI tools that never make it to production. Here's the framework that cuts through the noise and focuses on what actually ships.
- [The Agent Transformation Blueprint Most Enterprises Skip | Matthew Kruczek](https://matthewkruczek.ai/blog/company-transformation-agents.html): Moving from AI experiments to enterprise-wide transformation requires a new operating model. Here's a practical blueprint for transforming your company with AI agents.
- [Your AI Is Only as Good as Your Data Strategy | Matthew Kruczek](https://matthewkruczek.ai/blog/data-lifeblood-of-ai.html): Everyone says data is the lifeblood of AI, but few explain what that means in practice. Here's what enterprise leaders need to know about data strategy for AI success.
- [How Synthetic Data Unlocks AI When Real Data Is Off-Limits | Matthew Kruczek](https://matthewkruczek.ai/blog/synthetic-data-microsoft-ai.html): Synthetic data solves the AI data scarcity problem—especially when real data has privacy constraints. Learn how Microsoft technologies simplify synthetic data generation.
- [Azure AI Foundry vs Copilot Studio: Which One Wins? | Matthew Kruczek](https://matthewkruczek.ai/blog/microsoft-ai-studios-comparison.html): Confused by Microsoft's AI product portfolio? Here's a clear comparison of Azure AI Foundry, Copilot Studio, and Azure OpenAI Studio—and when to use each.
- [C# vs Java for Enterprise AI: Which Language Wins in 2026? | Matthew Kruczek](https://matthewkruczek.ai/blog/csharp-java-enterprise-ai.html): C# and Java both excel for enterprise AI, but they're not equal. Explore their strengths, tradeoffs, and when to choose each for production-grade AI systems.