Why 95% of Enterprise Agentic Projects Fail

AI is writing more code than anyone expected. Some of it is great. A lot of it is just okay. In this episode, Danny Thompson and Leon sit down with Matt DeBergalis, CEO of Apollo GraphQL, to unpack what it will take to move from a gold rush of mediocrity to production-grade agentic experiences that users can trust.

Guest Co-Host: Matt DeBergalis

https://www.linkedin.com/in/debergalis/

https://www.apollographql.com/

@ApolloGraphQL

SITE https://www.programmingpodcast.com/

Stay in Touch:

📧 Have ideas or questions for the show? Or are you a business that wants to talk business?

Email us at dannyandleonspodcast@gmail.com!

Danny Thompson

https://x.com/DThompsonDev

https://www.linkedin.com/in/DThompsonDev

www.DThompsonDev.com

Leon Noel

https://x.com/leonnoel

https://www.linkedin.com/in/leonnoel/

https://100devs.org/

📧 Have ideas or questions for the show? Or are you a business that wants to talk business?

Email us at dannyandleonspodcast@gmail.com!

We dig into the real gap behind AI project failures and it is not the models. Matt explains why agentic development stalls inside enterprises, how microservice sprawl blocks useful AI, and where GraphQL functions as the control plane that unifies data, streaming, and context so agents can actually do work. We cover the early hype around MCP servers, why many of them ship without OAuth, and a concrete checklist for securing costs and credentials before you flip the switch.

You will hear where shopping, search, and SEO are headed as prompt boxes replace search boxes. We get into the gravity that pulls models toward stacks with the most public code, what that means for React, Rust, Python, and the long tail, and how developers can future proof their careers by mastering fundamentals like orchestration, context control, and system design instead of chasing every weekly model benchmark.

We wrap with a practical path for job seekers. Breadth over tool loyalty. Weekly small projects. Use AI for the first 75 percent, then own the last 25 percent with clear prompts and better workflows.

Who this episode is for

- Engineering leaders trying to turn AI prototypes into products

- Senior and staff engineers learning agent orchestration

- Devs curious about MCP, GraphOS, and secure tool calling

You will learn

- Why 95 percent of agentic projects fail and what capability is missing

- How GraphQL unifies fragmented systems for agents, including streaming and precise context selection

- A security and cost control checklist for MCP style tool calling

- How hiring rubrics are shifting toward communication, systems thinking, and curiosity

- A weekly practice plan to build portfolio proof fast

Highlights

Gold rush of mediocrity and what to do about it

From REST to stateful agents and why the old web stack creaks

Every search box becomes a prompt box

The 75 and 25 rule for productive AI assisted coding

Tool breadth over tool loyalty for career advantage

Chapters

00:00 Cold open. Why most agentic projects fail

01:00 Theme setup. The gold rush of mediocrity

01:30 Host and guest introductions

03:00 MCP excitement vs reality. From laptop tools to real products

06:15 Security and spend. OAuth gaps, scoped keys, rate limits, audit logs

09:00 Distribution shift. Generative SEO and agentic checkout

13:10 Centralization gravity. Why models favor stacks with more public code

18:00 Foundations. Unifying services with GraphQL and streaming tokens

24:10 Controlling the context window with field selection

26:30 Should developers learn this now

31:30 Fundamentals over benchmarks. MCP, RAG, evals

42:00 Hiring in the agent era. Communication, systems thinking, curiosity

48:00 Prompt quality and the last mile

53:00 Audience question. Tools to explore and a weekly practice plan

59:30 Closing recap and CTA