Transcript
JOSH: It's Friday, November 13. This is Build or Be Replaced, powered by ScanBrief.dev. I'm Josh, here with Erik Anderson.
ERIK: Today's theme is access. Agents, chips, even Microsoft. Whoever can't see what's actually running loses.
JOSH: Stick around, Erik's got an AI pro tip at the end about making your agent prove its diagnosis before it touches production.
[pause]
JOSH: Headlines first. Microsoft Outlook and 365 are still down after twelve hours.
ERIK: Twelve hours. That headline is all I've got, and it's enough. If your whole company runs on one vendor's mail, you've got a single point of failure with a subscription fee.
JOSH: Do you run your own monitoring for that stuff?
ERIK: Always. I don't wait for a status page to tell me what I already know. Which brings up the next one.
JOSH: Uptime Kuma is trending on GitHub again.
ERIK: A fancy self-hosted monitoring tool. That's the whole headline, and the word that matters is self-hosted. I run it in the lab. It tells me something's broken before anyone's vendor admits it.
JOSH: And there's something called Edge0-145B. Frontier AI capabilities on your own computer.
ERIK: That's the claim in the title, so I'd want to see it run before I believe it. But the direction is right. Models you can run yourself keep getting more interesting. Self-hosted is the theme today.
[pause]
JOSH: Let's start with a paper. Claude Code debugging production databases.
ERIK: This one's a field observation out of ArXiv. The title is "Where LLM Agents Get Stuck Debugging Databases: Access More Than Reasoning." They watched one agent, Claude Code, diagnose eight database incidents in a production message-dispatch engine.
JOSH: And how'd it do?
ERIK: Here's the part that stings. In every incident where the root cause depended on runtime concurrency, the agent announced a cause, shipped a fix, and later walked it back. Every single one.
JOSH: Wait, really? Every one?
ERIK: Every incident of that type, per the paper. And each wrong guess cost a full deploy-and-observe cycle. So it's not just being wrong. It's being wrong in production, on a clock.
[beat]
JOSH: So the model just isn't smart enough?
ERIK: The title says otherwise. Access more than reasoning. My read is the agent was making confident guesses about a system it could only see through code and whatever logs it had. Concurrency bugs live in timing. You can't read timing out of source.
JOSH: How does that compare to what you do?
ERIK: PrimeBus was built around this problem. When a test fails, it spins up a Claude session to fix it. But the fix doesn't go straight to prod. It goes through tests, then Gandalf review, then the merge gate. The agent's confidence counts for nothing. Evidence counts.
JOSH: So the guardrails exist because the agent will be confidently wrong.
ERIK: Yes. I don't take that personally and neither should you. A human engineer does the same thing at 2 a.m. The difference is a human feels the shame and an agent ships the next deploy.
JOSH: What should people take from the paper?
ERIK: Give the agent read access to live runtime data before you ask it for a root cause. And make it show you what it observed, not what it believes. That's the tip, more or less. I'll finish it in a minute.
[pause]
JOSH: Big news. Jensen Huang has resigned from Nvidia.
ERIK: Co-founder, longtime CEO, public face of the whole GPU and AI acceleration run. The source calls it a shockwave through AI hardware and semiconductors, and that's fair. Nobody expected this.
JOSH: And the open source crowd is already reacting.
ERIK: The Hugging Face CEO pledged to keep fighting for open source AI after Huang's exit. TechCrunch notes it raises questions about Nvidia's $12.9 billion acquisition. I'm not going to guess the details beyond that, because the item doesn't give them.
JOSH: Why would Hugging Face say that now?
ERIK: Because a leadership change is when priorities get reopened. Whoever was promising things to the open source world isn't in the chair anymore. When you build on someone else's platform, their promises are only as good as their org chart.
JOSH: Does this change how you build?
ERIK: It confirms what I already do. PrimeRouter sits in front of every model call in my fleet. If one vendor changes direction, I change a chain in a config file, not a codebase. You don't need to predict who runs Nvidia. You need a router so you don't care.
[beat]
JOSH: There's a second Nvidia story, too. Chip tracking.
ERIK: MIT Technology Review reports Nvidia privately told Representative Bill Foster it can technically embed tracking in its chips. That's to enforce the anti-smuggling provisions in the Chip Security Act.
JOSH: But the industry said that wasn't feasible.
ERIK: Publicly, yes. Privately, Nvidia said it can be done. So the argument shifts from "can we" to "will we." For anyone doing infrastructure, it means hardware that reports on itself is coming. Plan for it. I'd rather inventory my own gear before a regulation makes me.
[pause]
JOSH: Last story. Amplitude and OpenAI. Shipping product improvements in one to two days.
ERIK: Amplitude built a product agent called Wave. It uses OpenAI models to identify high-impact feature ideas, hands the implementation to coding agents, and then measures the results automatically. The company says that gets improvements out in one to two days.
JOSH: That's fast. Is it just a coding agent?
ERIK: No, and that's why I picked it. The coding is the cheap part now. The interesting piece is the full loop. Idea, build, measure, repeat. Most teams automate the middle and leave a human to read the dashboard a week later.
JOSH: So the measuring is the real change.
ERIK: Yes. It adds accountability. If a feature ships and nobody checks if it worked, you didn't ship a feature, you shipped a guess. Same idea as PrimeBus. A fix isn't done when it merges. It's done when the tests prove it.
JOSH: Any caveat?
ERIK: It's a vendor case study, so it's the best story they've got. Still, ScanBrief scored ninety-six items across fifty-six sources this morning, and the ones that matter all share a pattern. Loops with feedback beat one-shot prompts.
[pause]
ERIK: This episode is sponsored by Prime Automation Solutions. If you're still doing it manually, we automate it. Also, special on a website — $250. primeautomationsolutions.com
[pause]
JOSH: Alright, what's the AI pro tip today?
ERIK: Make your agent prove its diagnosis before it ships a fix. When you ask an agent why something broke, add one rule to the prompt. It has to state its suspected cause, then name one check that would show it's wrong, and run that check first. A query, a log filter, a repro. The result goes in the output.
Here's why. That paper showed an agent announcing causes, shipping fixes, and walking them back. Each one cost a deploy. A falsifiable check costs thirty seconds. It's especially useful on anything involving timing, queues, or concurrency, where the code looks fine and the bug isn't in the code.
If the agent can't run the check because it lacks access, you've learned what to fix. Give it read-only access to the runtime data, not write access to prod.
That's your tip. Use it.
[pause]
JOSH: Binge all five episodes this weekend plus our YouTube shorts. Links at buildorbereplaced.dev.
JOSH: One more thing — we started a Discord for builders. If you're shipping AI, automation, or anything that makes a human obsolete — come hang out. Link at buildorbereplaced.dev.
ERIK: Post what you built. We'll post what we're building. Real wins, real builds, no fluff.
[pause]
ERIK: Build or be replaced.
JOSH: If you want these signals in your inbox every morning, scanbrief.dev. See you tomorrow.