Digital Velocity

There are a handful of genuine practitioner observations - routing work to cheap vs. expensive agents by task complexity, building unified work surfaces to eliminate app-switching, and the 'predictions came true sideways' reframe - but these are diluted by extensive conversational filler, mutual agreement loops, and surface-level recapping of widely-known AI news. The insight-per-minute ratio is low.

“Cheap agent for easy work, more expensive agent for harder work. You know, and the more expensive typically those are longer projects”

“I build a unified display system, funnel everything into one screen, and manage everything right there. And almost never go into the apps anymore.”

The 'predictions came true sideways' meta-framing is a genuine and interesting reframe, but most ideas in the episode are borrowed from others (explicitly crediting Jordan Wilson's 'human duct tape,' Paul Roetzer's agent-valuation model) or are widely circulating takes on token costs and tool consolidation. No contrarian or first-principles argument is developed at length.

“the predictions were about AI doing things, and the reality is about how us humans do things differently with AI”

“one of the things he he likes, the term he likes to use is human duct tape, and that's what it that's what it feels like”

Both hosts are working practitioners with real client engagements - Pat is onboarding enterprise healthcare and higher-ed clients on agent workflows; Erik is building LLM-driven analytics pipelines for e-commerce clients - giving the episode genuine ground-level credibility. However, neither operates at notable scale, there is no external guest, and their expertise is that of capable mid-market consultants rather than operators who have built or led these systems at a company of consequence.

“I have agents that just do stuff for me, decks and proposals. I wrote a bunch to start, I think eight to 10, and they're solid examples 'cause the client signed them”

“I'm working with clients now on more sophisticated stuff and that actually comes up like model selection. What are we gonna build this for?”

The episode name-drops real tools, platforms, and commentators (Claude CoWork, Codex, Gemini Enterprise, Snowflake, Jordan Wilson, Paul Roetzer), and includes one concrete client case with a governance policy and Gemini Enterprise purchase. However, statistics are consistently hedged or wrong ('I can't remember if this was past quarter reporting earnings'), clients are anonymized, and key claims about costs and adoption lack any hard data.

“if I remember a stat from I think a couple of days ago, that ChatGPT had, I can't remember if this was past quarter reporting earnings, but like, I don't know, negative like forty billion dollars in losses”

“I don't even remember the name of a corporation that burned through half a billion dollars in a month”

The episode is a collegial two-host chat with minimal structure and almost no productive friction; the dominant dynamic is mutual agreement punctuated by 'yeah' and 'hell yeah.' There is one sharp follow-up ('I'm curious, why aren't you setting that up as a system instruction?') and one reasonable open question about agent adoption maturity, but neither is pressed to extract a deeper or more specific answer.

“I'm curious, why aren't you setting that up as a system instruction?”

“Are we really in the age of agents yet? Are we like a little bit ahead of the curve and it's not quite there?”