Five separate conversations in one week stopped arguing about models and started arguing about the scaffolding wrapped around them. A regulated enterprise vendor, a chief AI officer at a cloud provider, a product leader at a construction software company and two solo builders - none of them talking to each other, all of them arriving at the same architecture.

💡 Top tip: before you optimise a prompt, map which paths in your system must never be probabilistic - then build those without a model in them at all.

Our ever-growing library of podcast summaries lives here - a quick refresher, an episode preview, or a way to get up to speed fast.

Featured in this round-up:

🕒 Estimated reading time: 13 mins. Time saved: 231 mins! 🔥

Not your topic this week? Try these five instead:

———————————————————————————

Five conversations, none of them about the same nominal subject, and the disagreements land in almost the same place every time. The sharpest is the interface. Rubrik's CPO described a vision where "you can operate all of rubric through this interface that you do not need the UI and dashboards at all", and the SaaStr host cheered it on with "I just, I can't learn another UX. I'm just done." Meanwhile Matt Wolfe spent months building a tabbed dashboard on Marketing Against The Grain, and the Blotato demo in the same show's other episode centres on a management screen where state stays synced both ways. Two positions, both shipping, pointing in opposite directions.

The second split is about where you put the person. Rubrik's answer is upstream approval of the plan. The Marketing Against The Grain answer is downstream inspection of the work, with the hosts coining "AI slob" for anyone who skips it. Tyler Folkman on The Growth Podcast is the honest middle: "The hard part of the loop is going to be validating that it passed the gate."

Then the convergence nobody planned. A regulated enterprise vendor and a solo creator, with nothing else in common, independently decided the dependable parts of their systems should not run on a model.

Key takeaways:

Chaitan Gupta, chief AI officer at Rackspace, put the argument in his episode title on Practical AI, and then spent the conversation refusing to name a favourite model. His framing is that enterprises face three questions before capability even enters the picture: local model or frontier model, what it costs, and whether sending data out means "your data sovereignty is not guaranteed". He quotes Satya Nadella on the consequence - "you're losing your alpha, right? That's your IP." The model is now a component you select against constraints, not a bet you place.

The builders demoing live systems said the same thing with their hands rather than their mouths. Rubrik's CPO on the SaaStr Podcast described a deliberate architectural fork: "We didn't wanna say, okay, we're gonna custom build all of these different workflows and anticipate the workflows that our customers are gonna have. We are going to build a platform such that it's generative." Their agent, Ruby, started life as a plain retrieval application over product documentation and became something that does forward-looking capacity planning - work that "would typically take them at least a day, if not longer".

Matt Wolfe's account of building his research dashboard on Marketing Against The Grain is the solo version of the same discipline. "My original prompt was me just like brain dumping via audio everything I wanted into this thing," he says, and then the real work started: "once the bones of the whole thing were built out, I just started going module by module and optimizing each module until they were all working." This week is Pod Shots #141 - It's Never The Model arriving independently from five directions at once.

Key takeaways:

The most transferable finding of the week came from two people who will never meet. Rubrik's CPO, explaining how Ruby handles cyber recovery, drew a hard line through the system: "we're using all the LLMs to help come up with the plans, but the plans themselves of how you're gonna do a recovery in a cyber scenario are deterministic and then are explainable to the customer." The reasoning is that "in a recovery scenario, you don't wanna be guessing and you don't wanna be using a probabilistic mechanism for recovery".

Matt Wolfe, building for an audience of one, made the identical call. His dashboard "out of the box does not actually need AI to track all of this stuff. It will just track it using like cron jobs and things like that." The AI curation layer is optional, sits behind a key you plug in, and can point at a local model if you have the hardware. He also still types his own social stats in by hand.

Where they split: they don't, and that is the point worth noticing. A regulated enterprise vendor with Fortune 2000 customers and a solo creator with a YouTube channel independently pulled the model off the dependable path and left it doing planning, drafting and curation at the edges. Chaitan Gupta supplies the theory: AI capability "is jagged", strong at code and merely adequate at email, so the architecture has to encode where you trust it. The default instinct to route everything through a model is the thing all three of them designed against.

Key takeaways:

Tyler Folkman, chief AI officer and head of product at JobNimbus, has the clearest working definition of a loop anyone has offered publicly. It is not a scheduled prompt. "Looping kind of assumes that you learn and it builds like this flywheel," he explains on The Growth Podcast. "Any good loop has this end that says, hey, AI, let's review what I did. Let's look at the log of all this back and forth and how do we take that and improve it?"

The distinction he draws is between a static skill and a self-improving one: "it's this idea of self-improving agents as opposed to a static skill. That the skill could just sit here and be the same forever or we could take this entire experience and feed it back into the skill and have the AI improve it." He demonstrated it live, spinning up a research agent and three prototype variants while still mid-conversation, then closing the loop with a learning hook that rewrites the skill from the session log.

The obvious risk is a system that quietly degrades itself, and his answer is unglamorous: Git. "Very normal for engineers, a little less normal for product people, though I think that's changing." If your skills improve themselves, they need version history more than a static prompt ever did. Matt Wolfe's months of module-by-module refinement is the same loop run manually, which is a useful reminder that the dashboard you admire was not one shot.

Key takeaways:

Everyone this week put a person in the middle. Nobody agrees where. Rubrik's position is upstream and singular: "ideally where we want it to be is that the human in the loop is approving the plan. So they're approving saying, hey, this is what we're gonna do. Do you agree that this is the right thing to do to achieve your outcome?" One decision point, before anything executes, at the level of intent.

The Marketing Against The Grain position runs the other way. The hosts have a name for the failure mode - "You don't read the code it's writing for you. You're just like asking it some questions and just taking it that's doing everything perfectly and just keep going. You're an AI slob." The Blotato founder agrees from experience: "I'm only able to trust it because I spent so much time reading it, understanding it, understanding what permissions I'm giving it." Her practical version is secret-shopping your own funnel, walking the path a customer would walk rather than assuming the automation is right.

Where they split: Rubrik gates intent, Marketing Against The Grain gates execution, and both call theirs the responsible version. They are catching different failures - plan approval stops you doing the wrong thing beautifully, output inspection stops you doing the right thing badly. Tyler Folkman names the part everyone else glosses: "The hard part of the loop is going to be validating that it passed the gate." A gate you cannot cheaply evaluate is decoration. His mitigation is a two-way door - a feature flag, a real customer, a fast reversal.

Key takeaways:

There is a failure mode in self-improving systems that only shows up after a few weeks, and Tyler Folkman has hit it. AI, asked whether a skill should be extended, always says yes. "It's never not going to say that. And if you always accept what it says, it will become too much. The skill gets too opinionated, too much context, too many tokens, and it kind of dilutes the whole value."

The shape of the bloat is specific. "It tries to create like mega skills that have like 10 use cases into one," he says, adding that this is "completely not what, if you look at like the best practices, even from like anthropic themselves, is like not the way to do it. The AI just really seems to like that. Let's just solve all the problems with a given skill."

Chaitan Gupta's jaggedness argument is the same discipline one layer down. He is entertaining about where the sharp edges are: AI-written email is "very verbose and they're very cliched", and "humans can still do better at writing emails maybe, than AI can", while the same system writes genuinely good software. If capability is uneven, a skill that claims ten use cases is claiming competence it demonstrably does not have across all ten.

Key takeaways:

Rubrik is designing its interface out of existence. "Our vision here is that eventually, you can operate all of rubric through this interface that you do not need the UI and dashboards at all, that it will be a fully agentic experience," their CPO says. The SaaStr host's reaction explains the demand behind it: "I just, I can't learn another UX. I'm just done, I can't do it. So I love that. You're pulling all the data so that I don't have to have an analyst to figure this out."

Both Marketing Against The Grain episodes did the opposite. Matt Wolfe's system is a dashboard with tabs - industry intel, brand mentions, reminders, an audience tracker, a newsletter page, a task list. The Blotato demo is built around a management screen holding every automation in one place, with two-way sync: "you can actually edit stuff directly in here and it will sync to whatever Claude tells you, or you could tell Claude to update and it will be shown here."

Where they split: the split tracks who the operator is. Rubrik's user is an administrator escaping accidental complexity in a product they never wanted to learn, so removing the surface is a gift. Wolfe and the Blotato founder are the operators of their own systems, and they need state visible at a glance to spot when something has quietly stopped working. Both are right in their own reference class. The expensive mistake is importing the enterprise chat-only answer into a workflow where you are the one accountable for the output.

Key takeaways:

The most sobering exchange of the week was about liability, not technology. Rubrik's CPO owns it without qualification: "at the end of the day, we're still responsible. Even when a customer makes a mistake and something isn't working, they blame us. And rightfully so, they should. It's our responsibility to make it obvious what the right choices are." The optimistic half of that argument is that today "people do make mistakes because there's so much complexity in setting up our product", and "people are gonna make mistakes more than agents if we build our agents with the right guardrails around it".

The SaaStr host pushed on the part that does not resolve. Customers and partners will build things you never anticipated, and "you're gonna be responsible for everything. And you're not gonna predict what your customers are gonna do with it. You're not gonna know." Rubrik's own answer includes handing agent-building to systems integrators and partners operating on the customer's behalf, which widens the surface further.

Chaitan Gupta arrives at the same place through procurement rather than product. His customers now demand proof before deployment: "You got to prove that your model works on data that's relevant to me before you deploy it for my customers." Combined with the sovereignty concern, that is a market pricing in agent risk before regulators get there.

Key takeaways:

———————————————————————————

Three unresolved questions from this week worth carrying into your team.

Which of your paths must never be probabilistic. Rubrik and Matt Wolfe both mapped this deliberately and built those paths without AI, and most teams have never drawn the line at all. The exercise is cheap: list what your system does, mark what costs you a customer if it returns something plausible but wrong, and check what is currently running through a model on that list.

Where your human gate sits, and whether you can afford to check it. Rubrik gates the plan, Marketing Against The Grain gates the output, and Tyler Folkman's warning is that validating the gate is the hard part nobody budgets for. Pick your position deliberately rather than inheriting it, then staff the failure the other camp catches.

Whether the interface is the thing you remove or the thing you ship. Rubrik is deleting the dashboard while two solo builders made the dashboard the product, and the answer depends entirely on whether your user operates the system or escapes it. Get that wrong and you will build a beautiful chat interface for someone who needed to see state at a glance.

Want more of the same? Try these five:

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know what you liked or want changing! 🚀👋

Alastair 🍽️.