
Sign up to save your podcasts
Or


Two AI agent swarms took on giant coding jobs this year. Cursor's wrote a web browser its CEO said "kind of works." Anthropic's wrote a C compiler that could build the Linux kernel. Jerry Duff, VP of Technology at Critical Propulsion, explains what separates a swarm that does real work from one that just looks busy: feedback loops, test-driven development, who reviews the agents, what it all costs, and what he learned running his own team of Hermes agents over a weekend.
Full disclosure: Jerry's company sells this way of working.
AI disclosure: AI tools helped produce this episode (research, editing, captions and thumbnail art). The conversation and footage are real and unaltered.
Find Jerry: criticalpropulsion.com, and on LinkedIn
0:00 Two swarms, one giant job
1:28 Meet Jerry Duff
2:39 What Makes a Swarm Work
6:53 Proving the Work Is Done
11:36 Who Checks the Agents?
13:58 Keeping Humans in the Loop
18:16 The Cost of Running AI
23:34 From Client to Builder
26:39 Good Enough to Ship
29:30 The Review Bottleneck
31:03 The Price of Speed
36:00 Inside the Agent Lab
Watch on YouTube: https://youtu.be/5W1FC-dJo5M
More from The Arqiv: https://arqiv.fm
John Capobianco, Head of AI at Itential, turned a real computer network into a 3D world you can fly through like a drone, using MCP servers for Blender and Unreal Engine 5.
We also get into rebuilding the classic Battle Chess with Veo 3 and Nano Banana, training a digital fruit fly to monitor network health, why "low" security vulnerabilities matter more in the AI era, onboarding AI agents like junior engineers, canary prompting for context rot, cutting token costs, and whether new engineers still need the fundamentals.
Itential: https://www.itential.com
Catch John at AutoCon 6 in Tucson, Nov 16-20.
Watch the full video on YouTube: https://youtu.be/0U5_VKiQNJE
Chapters
00:00 Cold open
00:58 Building an intro with AI agents
01:44 Battle Chess rebuilt with Veo 3
03:04 Video models vs world models
03:52 Blender MCP: photo to 3D
04:22 Fly through a network in Unreal Engine 5
05:22 Google Genie and World Labs
07:26 AI-driven NPCs in games
08:47 A fruit fly that monitors networks
10:41 Meet John and Itential
12:46 Why agents start read-only
14:07 Why low CVEs are now high risk
16:27 Reasoning plus deterministic workflows
21:41 Engineers as managers of agents
25:03 IT becomes HR for agents
28:22 A cheap go/no-go decision agent
31:44 Canary prompting and context rot
33:15 Staying current and always-on assistants
36:01 Will platforms block AI agents?
38:32 Cutting token costs
41:47 Preparing students for AI
44:41 800 pages vs 8 lines of JSON
49:25 Learning how to learn
50:31 Where to find John
The ARQIV: Auditing Reality. Better questions. Better understanding.
https://arqiv.fm
Instagram | TikTok | LinkedIn
Who checks AI agents before they act on their own? Red teamers: people and tools that attack an agent on purpose, like a real adversary would, to find the weak spots before customers or attackers do.
In April, an AI coding agent at PocketOS found an over-powered password and deleted a live production database, backups included, in nine seconds. Now AI writes about 75% of Google's new code, and agents are starting to click, email, move money and change files without a human in the loop. So who's checking them?
Steve Ryan sits down with Pramin Pradeep, co-founder and CEO of BotGauge, which red-teams AI agents before they reach production. We cover why agents go wrong when they're bolted onto old systems, how prompt injection beat a $100 refund guardrail, why "always allow" is a trap, who's accountable when an agent fails, and what students should practice for an AI-run workforce.
CHAPTERS
0:00 Who checks AI agents?
1:22 Meet BotGauge
4:54 The 9-second database deletion
7:58 Red teaming AI agents
10:19 Prompt injection
13:26 Humans in the loop
16:45 Who's responsible when AI fails?
22:26 The human bottleneck
26:09 Agents vs. automation
28:33 AI writes the code
30:01 When do agents go autonomous?
30:44 Jobs and the next workforce
GUEST
Pramin Pradeep, Co-founder and CEO, BotGauge
https://www.botgauge.com
Find Pramin on LinkedIn: Pramin Pradeep
SOURCES
PocketOS deletion (Decrypt): https://decrypt.co/365897/ai-agent-deletes-startup-database-9-seconds-founder-says
Replit / SaaStr database wipe (eWeek): https://www.eweek.com/news/replit-ai-coding-assistant-failure/
Google: 75% of new code AI-generated (Semafor): https://www.semafor.com/article/04/24/2026/google-ceo-says-75-of-companys-new-code-is-ai-generated
Veracode GenAI code security (SD Times): https://sdtimes.com/agentic-security/veracode-finds-ai-generated-code-security-has-barely-improved-since-last-year/
OWASP Top 10 for Agentic Applications: https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications/
PHP2Shell fact check (Al Jazeera): https://www.aljazeera.com/news/2026/10/22/whats-the-php-coding-flaw-that-could-pull-down-70-percent-of-the-internet
Also cited: Gartner agentic AI forecasts (June 2025), McKinsey State of AI 2025, WEF Future of Jobs 2025, Stanford Digital Economy Lab "Canaries in the Coal Mine" (Aug 2025), Moffatt v. Air Canada (2024 BCCRT 149). BotGauge figures are company-reported.
THE ARQIV
Auditing Reality. Ask. Research. Question. Inspect. Verify.
https://arqiv.fm
More than 6,000 planets confirmed beyond our solar system. More than 60 years of listening to the sky. And still, silence. Is it because nobody's out there⦠or because almost nobody ever got the fuel to call?
Dr. Terry Adair, oral and maxillofacial radiologist, six-time Ironman finisher, and author of "Where Are the Little Green Men? The Adair Hypothesis," joins The ARQIV with a new answer to the Fermi Paradox: the bottleneck isn't intelligence. It's energy.
In this episode:
Terry's books (Where Are the Little Green Men? and Earth: The Goldilocks Planet): https://terryadair.com
Coming next from Terry: The Road to Consciousness.
Chapters
00:00 6,000+ planets and 60 years of silence
00:46 Meet Dr. Terry Adair
02:58 What an oral & maxillofacial radiologist does
07:40 Starting over at 55 (and an Ironman at 50)
10:02 The Fermi Paradox explained
15:45 The Adair Hypothesis: it's about energy
19:53 Gasoline and dense, stored energy
21:14 The 93-million-mile coal train
23:59 Two pathways: planetary and civilizational energy
31:38 Why a smart octopus can't build a radio
36:18 The Great Silence: a privilege, not loneliness
38:41 Can humanity skip a rung on the Kardashev scale?
40:48 Dyson spheres, swarms and solar power in space
42:56 Entropy: the cosmic pickpocket
45:51 Lightning round
47:32 Where to find Terry's books
Watch the full video on YouTube: https://www.youtube.com/@thearqiv
More at https://arqiv.fm
The ARQIV: Auditing Reality. Better questions. Better understanding.
AI agents can now write code, open pull requests and act on your behalf. So who checks their work? Steve Ryan sits down with Matt Davis, co-founder of Critical Propulsion ("Humans lead. Agents execute."), to work out what agents should decide on their own, where a human approval gate belongs, and why rolling out agents is a change-management problem, not just a tech one. It starts with a real story: Meta's Muse AI gave a stranger a YouTuber's home address on Facebook Marketplace.
π€ This episode was edited with the help of AI tools, but it was made by a real person: a real conversation between real people, edited and reviewed by Steve.
Key takeaways
Chapters
0:00 Cold open
0:25 Where does an agent's permission end?
2:08 Gates at every function
3:23 YOLO mode vs. guardrails
6:00 Where do humans fit?
6:49 The human bottleneck
8:17 Verification & hallucinations
9:30 It's change management, not tech
11:03 Lessons from Amazon Robotics
12:01 Reskilling: judge the outputs
13:28 One person, many agents
15:39 Keeping up with the models
17:48 The cost of tokens
21:13 Open & Chinese models
22:11 Is AI coming home?
24:17 Security, risk & sandboxes
27:46 Why AI isn't plug-and-play
29:04 Inside Critical Propulsion
31:42 Knowledge workers aren't dead
Stories mentioned
Meta launches Muse Wonder (Android Authority)
Muse shared a YouTuber's home address (Dexerto)
Claude Opus 5.5 (The Batch)
Anthropic files confidential IPO paperwork (Yahoo Finance)
Why residents say no to AI data centers (NPR)
Nvidia RTX Spark for local AI (The Decoder)
OpenAI publishes 377 math results (Interesting Engineering)
Watch the full video on YouTube: youtu.be/eIONRMEX0CQ
ARQIV β Auditing Reality. Ask. Research. Question. Inspect. Verify. arqiv.fm
The launch of the new SIRI AI has me asking some questions about what the future of the phone UI. How can AI replace Apps like games and other apps.
Is AI rotting our brains, or is it the ultimate learning tool? MIT scientists hooked 54 people up to EEG sensors and found those using ChatGPT had the lowest cognitive engagement. But that's only half the story.
In this documentary-style deep dive, we explore the massive gap between the panic surrounding artificial intelligence and what the data actually proves. While a Turkish study showed that students using regular ChatGPT scored 17% worse on exams, those using a hint-only AI performed 127% better during practice. The crisis isn't the technology; it's the design. We also audit reality on the job market, uncovering why AI-exposed jobs for 22 to 25-year-olds have dropped by 16%, completely hollowing out the traditional entry-level ladder.
If you want to understand how tech is shaping our future without the heavy jargon, subscribe to ARQIV so you don't miss our next deep dive into the reality of tomorrow's economy.
How is AI changing mental health treatment? Lumos AI and Headlamp Health are using artificial intelligence and machine learning to revolutionize AI drug discovery and aim to deliver personalized mental health care in just minutes.
For decades, psychiatric treatment for conditions like major depressive disorder has relied on a frustrating, years-long trial-and-error process. Erwin Estegariba explains how AI is finally bypassing this guesswork. By analyzing complex medical histories, Lumos AI creates a complete "moving picture" of a patient's brain chemistry to accurately predict medication successβstriving to move past the 50% "coin flip" of traditional psychiatry.
Disclaimer: The information provided in this video and description is for educational and informational purposes only and does not constitute medical advice. Always consult with a qualified healthcare professional regarding any medical conditions or treatment decisions.
0:00 - The AI Platform Transforming Mental Health1:18 - Why Psychiatric Drugs Keep Failing3:55 - Creating a "Moving Picture" of Your Brain5:47 - The Blood-Brain Barrier Problem7:27 - How Lumos AI Actually Works11:08 - The 50% "Coin Flip" of Psychiatric Trials13:21 - Reconstructing Medical History in 8 Minutes17:22 - Rescuing 30 Years of Abandoned Medicine21:01 - Why All Depression Isn't the Same22:53 - Why You Shouldn't Use ChatGPT for Therapy27:04 - Shrinking a 7-Year Diagnosis to 7 Minutes29:02 - The Truth About Prescribing Multiple Medications
Every week there's another trillion-dollar AI announcement. Data centers, energy grids,
digital twins of the planet. But all of it has to be physically built β with materials,
supply chains, and electricity that were already breaking before AI showed up.
Josh Dorfman co-founded Planted with two former SpaceX engineers to make carbon-negative
building panels out of grass, and just landed a 10-million-panel contract with the largest
home builder in America. Before that, he ran Vine.com inside Amazon. Now he runs Supercool,
a media platform tracking which climate technologies are actually scaling and which are
well-funded slide decks.
This conversation goes into the gap between what AI promises and what physical infrastructure
can actually deliver β and ends somewhere neither of us expected.
βββββββββββββββββββββββββββββ
CHAPTERS
00:00 The question nobody wants to answer
02:30 Why data centers are a materials problem, not just an energy problem
05:00 "AI is the enabler of cheap renewable energy" (the optimistic case)
08:30 The Lords of AI and the gap between slide decks and market
11:30 BrainBox, Natrix, and what AI is actually doing in climate right now
18:00 Planted: how you make a building panel out of grass
22:00 The carbon math β what "carbon negative" actually means
27:00 The SpaceX playbook (and a $140 pickle-barrel mixer)
31:30 Why silos killed projects at Amazon
42:00 The Trove rule: "Nobody talks about climate"
48:00 Is AI's buildout ultimately good or bad for the climate?
51:00 How to tell real climate tech from vaporware
54:30 Where Steve pushes back: the case against more data centers
βββββββββββββββββββββββββββββ
ABOUT ARQIV
ARQIV is a science, technology, and systems documentary channel hosted by Steve Ryan.
Twelve years at Amazon, four as a Senior Product Manager in Amazon Robotics. Now auditing
reality one system at a time.
π arqiv.fm
π¬ Newsletter: [link]
π Podcast: [link]
βββββββββββββββββββββββββββββ
RELATED ARQIV EPISODES
β NVIDIA's Earth-2 [link]
β Project Genesis [link]
β [Hagan episode when published]
βββββββββββββββββββββββββββββ
ABOUT JOSH DORFMAN
Josh is co-founder of Planted and host of Supercool. Find him at [supercool link] and
[planted link].
#AI #DataCenters #ClimateTech #Infrastructure #Supercool
Discover how non-invasive Brain-Computer Interfaces (BCI) are making mind-reading a reality without surgery. Learn how neurotechnology is changing how we interact with the world forever.
Imagine texting, emailing, or searching the web just by thinkingβno Neuralink surgery required. In this deep dive, we explore the breakthrough science of non-invasive BCIs, the recent Nature Neuroscience papers proving brain-to-text is possible today, and the closing gap between "difficult" and "practical" technology. We also tackle the critical debate over "Neural Rights" and how to protect the last truly private space you have: your mind.
If you want to understand how tech is shaping our future without the heavy jargon, subscribe to ARQIV so you don't miss our next deep dive into the technology of tomorrow.