Real AI Agents vs. Fake Ones (The 30-Second Test) | Practical AI Ep 31

Practical AI: Episode 31

Real AI Agents vs. Fake Ones (The 30-Second Test) 

Watch on YouTube

Published: March 7, 2026

TL;DR

This week’s episode unpacks the Anthropic vs. Pentagon drama that saw the company labeled a “supply chain risk” while OpenAI swooped in with a replacement contract, the Cal AI acquisition by MyFitnessPal proving distribution beats code every time, a Supreme Court ruling cementing that pure AI-generated content cannot be copyrighted, and OpenAI’s record $110 billion funding round. Plus: a live framework vibe coding demo and Olga’s 30-second test to cut through “agent washing” and find real AI agents in your tools.

Table of Contents

About This Show

Practical AI is a weekly live show (Fridays 11am CT) hosted by Olga Pechnenko and Chris Pearson that cuts through AI hype to deliver news, trends, and hands-on tips for builders and founders. Unlike technical AI podcasts, Practical AI focuses on business applications and ROI: what actually works, what’s hype, and what you can implement Monday morning.

What You’ll Gain

  • Understand the Anthropic-Pentagon fallout and why being labeled a “supply chain risk” could crater enterprise sales for any AI company, plus how OpenAI capitalized on the chaos with a government deal just hours later.
  • Learn why distribution beats code through the Cal AI acquisition story, where teenage founders turned a simple photo-to-calories app into $30M ARR by cracking Gen Z distribution with influencers and a Mr. Beast sponsorship.
  • Discover the 30-second agent washing test to identify which “AI agents” in your existing tools actually do something useful versus which ones are glorified chatbots sending you to documentation.
  • See framework vibe coding in action as Chris live-demos adding an SCSS compiler to PageMotor using Claude Code, showing why structured frameworks produce sustainable results while naked vibe coding leads to 80% project abandonment.
  • Gain clarity on the AI copyright ruling and what it means for vibe-coded products: the Supreme Court confirmed AI-generated content without human input cannot be copyrighted, creating new legal uncertainty for creators.

Biggest Takeaway to Implement: Go test every AI assistant in every tool you already pay for. Ask it to do something with your actual data, not answer a generic question. If it gives you “wow” (pulls records, takes actions, knows your context), use it daily. If it gives you “disappoint” (sends documentation, gives generic advice), it’s a chatbot pretending to be an agent. You’re probably sitting on gold mines you’ve never explored.

PageMotor and Practical AI Updates

Free, informative, and FUN!

Frequently Asked Questions

What happened between Anthropic and the Pentagon?

Anthropic refused Pentagon demands for unrestricted use of Claude, specifically objecting to mass surveillance of Americans and fully autonomous weapons. The Pentagon responded by designating Anthropic a “supply chain risk” (a label typically reserved for foreign adversaries), forcing defense contractors to stop using Claude. OpenAI signed a replacement deal hours later. Read more below.

How did teenage founders sell their AI fitness app for tens of millions?

Cal AI, built by 19-year-olds using OpenAI’s API, hit 15 million downloads and $30M+ ARR in under two years. They cracked Gen Z distribution through influencer marketing and a $500K Mr. Beast sponsorship, then sold to MyFitnessPal. The lesson: anyone can vibe code an app, but distribution is the real moat. Read more below.

Can AI-generated content be copyrighted?

No. The Supreme Court declined to hear the Thaler v. Perlmutter case on March 2, 2026, cementing lower court rulings that copyright requires human authorship. Pure AI output cannot be protected. The question of how much human prompting qualifies as authorship remains legally murky. Read more below.

What is the 30-second AI agent test?

Open any “AI agent” or “AI assistant” in your software tools. Ask it to do something with your actual data (not a generic question). If it pulls your records, takes actions, and knows your business context, it’s a real agent. If it sends documentation or gives generic advice, it’s a chatbot with marketing spin. Read more below.

What AI agents actually passed the test?

Three verified winners: HubSpot Breeze (reads/writes CRM data, creates tasks, analyzes pipelines), Descript Underlord (edits video from prompts, removes filler words automatically), and Notion Agents. Many others claiming “AI agent” capabilities failed the data-access test. Read more below.

What is framework vibe coding vs. naked vibe coding?

Naked vibe coding means sitting down with AI and no pre-existing structure. It works for demos but 80% of projects don’t survive the first update. Framework vibe coding uses established patterns, skills, and guardrails so AI outputs fit into maintainable systems. Read more below.

How much did OpenAI raise in their latest funding round?

OpenAI closed $110 billion from Amazon ($50B), Nvidia ($30B), and SoftBank ($30B), valuing the company at $730 billion pre-money. For context, total US venture funding in 2023 was $170 billion. OpenAI raised 65% of that in a single week. Read more below.

Practical AI Episode 31: The 30-Second Agent Test

Key Definitions

What is Agent Washing?

Marketing spin where companies label basic chatbots as “AI agents” when they cannot actually access user data or take actions. A real agent works inside your existing tools, acts on your information, and does things autonomously. A washed “agent” sends you to documentation and gives generic advice you could find anywhere.

What is Framework Vibe Coding?

Using AI to generate code within established patterns, skills, and architectural guardrails. The framework provides highways for AI output to follow, producing maintainable results. Contrasts with “naked” vibe coding (no structure), where 80% of projects fail at the first update because there’s no foundation to build on.

What is a Supply Chain Risk Designation?

A Pentagon label typically reserved for foreign adversaries like Huawei, requiring all defense contractors to certify they don’t use the designated company’s products. Applied to Anthropic in March 2026, marking the first time an American AI company received this designation.

Quotable Moments

“80% of vibe-coded projects don’t survive the first update. Because it’s hard to update something when there’s no rules or whatever.” — Olga Pechnenko on why framework vibe coding beats naked approaches

“Claude Code is that coding partner, that fictional coding partner who’s never existed before. I find myself speaking to this thing in ways that you would absolutely speak to a human.” — Chris Pearson on the new reality of AI-assisted development

“If you don’t feel wow with an AI agent, it’s not an AI agent. If it just tells you where to look, if it’s not connecting dots for you, if it’s not doing stuff for you, it’s just a traffic cop.” — Olga Pechnenko on cutting through agent washing

“AI is a rocket ship. And the thing about traveling interstellar is that if you go to another star system, you can never go back to where you came from. We’re on that curve right now. There is no going back.” — Chris Pearson on the irreversibility of AI transformation

1:51 Anthropic vs. OpenAI: Government Drama Explodes

Key Stat: Fortune 100 Government Exposure

About 70% of Fortune 100 companies have government contracts. Anthropic’s target market for its planned IPO now faces a major barrier if the supply chain risk designation sticks.

Anthropic refused Pentagon demands for unrestricted Claude access, objecting to mass surveillance and autonomous weapons. The Pentagon responded by designating them a “supply chain risk” (typically reserved for foreign adversaries). OpenAI signed a replacement deal hours later. Dario Amodei’s leaked 1,600-word memo calling the OpenAI deal “80% safety theater” backfired badly. He later apologized.

Consumer response was striking: Claude surged to #1 in 20+ countries. But enterprise looks grimmer: Lockheed Martin and other defense contractors are already cutting ties.

9:08 Teens Sell $30M ARR AI Fitness App Overnight

Cal AI by the Numbers

MyFitnessPal acquired Cal AI after nearly a year of negotiations. The app hit 15 million downloads and $30M+ ARR in under two years, built by 19-year-old founders Zach Yadegari and Henry Langmack.

Cal AI’s concept was simple: point your camera at food, get calorie estimates. Built using OpenAI’s API with basic prompting, it succeeded through distribution, not technical superiority. The founders ran two apps (one for users, one for influencers with large text for video visibility), paid $500K for a Mr. Beast sponsorship, and held Sunday night standups while attending college.

The lesson: vibe coding is one thing, distribution is completely different. Anyone can build these apps now. The premium is on getting users.

13:34 Voice Coding Revolution: Talk Code While Driving

Claude Code Voice Mode Rollout

Anthropic launched voice mode for Claude Code on March 3rd, currently live for 5% of users. Type /voice to toggle, then speak commands like “refactor the authentication middleware.”

The paradigm shift is real: with voice input, you have to think upfront. It becomes like chess, planning before execution. Voice coding will likely start as a secondary modality (engineers in self-driving Teslas), but a new skill is emerging: extracting abstract mental models into spoken instructions.

17:39 OpenAI Building Its Own GitHub for AI Agents

OpenAI is reportedly building a GitHub rival, an internal code hosting platform designed for how AI codes. GitHub has been struggling with reliability as vibe coding explodes usage. The theory: AI doesn’t code line by line. It can update thousands of files automatically, and current Git infrastructure isn’t wired for that pattern. Chris was skeptical this is an architectural issue rather than an input/output capacity problem, but OpenAI clearly sees an opportunity in the tooling layer.

20:38 NotebookLM Levels Up With Video and Styles

Google’s NotebookLM released cinematic video generation for AI Ultra subscribers. The bigger update: new theme and style options (sketched infographics, kawaii, professional, scientific). The video feature creates slideshow-style animated clips from your notes, generating graphics relevant to the content. Free for anyone with a Gmail account.

24:23 Tech Giants Agree to Fund New Power Plants

On March 4th, major AI companies agreed to fund new power plants directly, shielding consumer electricity prices from AI infrastructure costs. The Trump administration pushed back on the traditional model where taxpayers foot infrastructure bills: you’re building for AI, you pay upfront.

26:13 Unions and Politicians Fight AI Job Takeover

An unlikely coalition (AFL-CIO, Steve Bannon, progressive Democrats, religious leaders) signed a “pro-human declaration” emphasizing human agency over AI-driven job replacement. This signals coming “human in the loop” regulations. Chris’s take: AI does some tasks better than humans. That repositions us, it doesn’t make us worthless. Legislating against AI efficiency feels like a losing battle.

28:18 Apple M5 MacBook Air: Local AI Just Got Serious

M5 MacBook Air Specs

Apple announced the M5 MacBook Air on March 3rd with 4x faster AI performance than M4 and 9.5x faster than M1. Starts at 512GB storage (double previous). Ships March 11th.

The M5 includes Neural Accelerators in every GPU core, enabling local LLM operation without cloud dependencies. Load Qwen, Llama, or other open source models locally, train on your circumstances, pay zero inference costs. Apple clearly noticed the Mac Mini craze (sold out for local AI) and responded. Chris’s prediction about edge devices is materializing.

The Ruling

On March 2nd, the Supreme Court declined to hear Thaler v. Perlmutter, cementing that copyright requires human authorship.

The ruling creates ambiguity for vibe-coded products. If you write all scripts and plan all dialogue, but AI generates visual output, where’s the line? The Copyright Office says prompting alone doesn’t constitute authorship. Olga’s practical concern: if your logo was AI-generated, you technically cannot copyright it. Every vibe-coder needs to document human involvement.

32:42 Live Vibe Coding Demo: Framework Magic in Action

Framework vs. Naked Stats

About 80% of naked vibe-coded projects don’t survive the first update. Without established patterns, AI outputs lack consistency.

Chris demonstrated live framework vibe coding: adding an SCSS compiler to PageMotor using Claude Code. Naked vibe coding (no plan, no infrastructure) produces unmaintainable outputs. Framework vibe coding uses established patterns so AI outputs fit sustainable systems.

The demo showed real-time troubleshooting: PHP version errors, CSS compilation issues. Claude Code traced the architecture, identified problems, proposed fixes. Something that would have taken “probably a couple days searching Stack Exchange” resolved during the show. Only one file needed to change for the compiler swap, a sign of good architecture. Chris’s revelation: Claude Code feels like the fictional coding partner developers always wanted.

41:57 The 30-Second Test: Spot Real AI Agents vs. Fakes

The Agent Washing Problem

Out of a dozen “AI agent” claims Olga tested, only three passed: HubSpot Breeze, Descript Underlord, and Notion Agents.

Olga’s breakthrough: when her AI suggested asking HubSpot’s Breeze about her ICP, she didn’t expect much. Breeze analyzed her CRM data, pulled closed deals from six months, delivered actual ICP analysis. “Wait, what? It actually used my data.”

The formula for a real agent: a tool you already use plus AI embedded natively that acts on your data. The 30-second test: open any AI feature, ask it to do something with your actual data. “Update my deals with no activity in 30 days.”

Wow signals (real agent): pulls your records, takes actions, knows your business context. Disappoint signals (chatbot): “I don’t have access to your data,” sends documentation, generic advice. If you don’t feel wow, it’s not an agent.

58:28 AI Funding Explosion: OpenAI’s $110B Mega Round

OpenAI’s Record Round

OpenAI raised $110 billion: Amazon $50B, Nvidia $30B, SoftBank $30B. Pre-money valuation: $730 billion. For context, total US venture funding in 2023 was $170 billion. OpenAI raised 65% of that in one week.

Excluding OpenAI, AI funding was $2 billion across 51 companies, healthy but not historic. The gap between OpenAI and everyone else isn’t closing; it’s becoming a canyon. Notable rounds: Ayar Labs ($500M for optical IO), Galaxy Bot ($364M for industrial robots), Shenzhi Technology ($321M for autonomous driving chips).

OpenAI committed to 5 gigawatts of Nvidia compute and expanded its AWS deal to $138 billion total. Infrastructure is the new moat.

1:09:20 Final Takeaways: Let Go and Let AI Do the Work

Chris’s theme: let go. Re-examine your workflow for areas where you’re holding on too tightly. Deploy AI in the areas that scare you. If Chris can become a vibe coder after resisting for 16 months, anything is possible.

Olga’s takeaway: test every AI assistant in every tool you use. Run the 30-second test. Use reverse prompting: “Based on everything you have about me, what questions should I be asking?”

The analogy that stuck: AI is a rocket ship to another star system. Once you go, you can never go back. The rocket has already launched.

Keep Learning