Code To Cloud

There are a handful of genuinely useful ideas - most notably that AI tooling is inflating PR queues without shrinking lead time, and that Claude Code hooks can run deterministic checks to autonomously remediate code smells before a commit. However, these insights are spread thin across a lot of casual back-and-forth repetition, mutual validation, and meandering anecdotes that dilute the signal.

“you're getting like more code being created. But then you have the reviews. The reviews are going up as well. And so to me, what that means is like your lead time is not changing. All you're doing is just creating more inventory.”

“the hook will have a set of rules and will look for things like high psychomatic complexity and look for things like long methods... And not only will it discover those in your code essentially as you're writing it... But then it will go and then fix it”

The framing of 'harness engineering' around agentic tool hooks and the idea of risk-based PR routing are moderately fresh angles, but the bulk of the conversation rehashes well-worn DevOps canon (trunk-based dev, progressive delivery, Jez Humble's decade-old CI certification story) simply applied to an AI context - which is a familiar genre of content in 2024-2025.

“harness engineering... the actual harness, so for example, cloud code itself and all the tooling that you can incorporate into that, whether it's like hooks where you can run deterministic checks”

“if you have a rule that says, okay, if this is a high-risk change, whatever you define high-risk change to be... you can route it based on that versus just everything goes through a PR queue”

Paul Duvall has genuine practitioner credentials - he authored a canonical book on Continuous Integration roughly 20 years ago, founded and sold Stelligent, worked at AWS, and is now running an AI consultancy - making him a legitimate operator rather than a career podcast guest. His depth on the topic is real, though his AI-specific work is more 'enthusiastic early adopter' than deep technical innovator.

“I wrote a book probably, I think it's like almost 20 years ago now, called Continuous Integration”

“I started a company called Stelligent and we focused on what became known as helping enterprises apply DevOps practices and ultimately on AWS specifically... built up that company and sold that in 2017”

The episode names real tools, books, and events (Claude Code hooks, Jez Humble, Accelerate, AI Engineer World's Fair at 6,000 attendees), but customer evidence is kept almost entirely abstract - no companies named, no before/after metrics, no dollar figures or timelines from actual engagements. The open-source repo with '25 or so patterns' is mentioned without enough detail to be actionable.

“I have been using Cloud Code for the past 13 months, basically since it was GA, I think it was back in may of last year”

“there's like six different event hooks you can use in Cloud Code”

The host is engaged and occasionally surfaces interesting follow-ups (e.g. asking how hook-based checking works in real-time), but he frequently rambles in his own questions, validates unchallenged ('that's a really good point'), and contributes long personal anecdotes that eat time without extracting more from the guest. There is no meaningful pushback or stress-testing of any claim.

“So, I mean, that's a really good point. I've never heard it being called harness engineering. Is that what you said?”

“I'm conscious of time so i know you're a busy guy um any words of wisdom or feedback”