Build or Replaced
Today: Dots: Always-on agents | September 2026: The world today, as seen by one Polish guy | Livenerf: Has Opus 5.5 been nerfed yet? Episode date: 2026-09-30.
Download MP3 →ERIK: ScanBrief scored 104 items across 56 sources before you even had coffee, and half of Hacker News is arguing about whether Anthropic quietly nerfed a model. Busy morning. JOSH: It's Wednesday, September 30th. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. JOSH: Stick around — Erik's got an AI pro tip at the end about picking the right model for the job instead of just reaching for the newest one. [pause] JOSH: Let's get into the headlines. First up — PS5 Relapse Exploit is trending on Hacker News. That's a console jailbreak thing, right? ERIK: Yeah, security researchers found a way back in on a system Sony thought was locked down. Every "permanent" patch has a shelf life. I think about that every time someone tells me a config is "locked." JOSH: Next — Backblaze dropped their Q2 2026 drive stats. ERIK: Backblaze publishes real failure rates across hundreds of thousands of drives. No vendor marketing, just numbers. I wish more companies published stuff like that about their own uptime. [beat] JOSH: And Vermont is swapping power plants for home batteries? ERIK: Distributed storage instead of centralized generation. Same pattern I use for compute — don't put everything through one choke point. Spread the risk. JOSH: One more — there's a thread going around about AI browser agents finally clicking buttons reliably instead of hallucinating them. ERIK: That's the boring engineering nobody claps for. Getting an agent to actually click the right pixel is harder than getting it to write a paragraph. Reliability is the whole game right now, not cleverness. [pause] JOSH: Okay, let's go deeper. First story — this "Dots" framework for always-on agents. What's the big deal? ERIK: Dots lets you deploy lightweight agents that just stay running and react to events in real time instead of you polling or triggering them by hand. That's not a new idea to me — that's Tuesday. JOSH: Wait, you're already doing that? ERIK: That's the whole Bobaverse. I've got twelve agents running across the fleet right now — Neo, Homer, Bill, Echo, Gandalf, a mix of Claude and GPT — and they're not waiting for me to prompt them. They're subscribed to events on PrimeBus and they act when something happens. JOSH: So when something breaks at 3 AM— ERIK: Something already tried to fix it before I woke up. That's the point of always-on. The second a framework like Dots makes that easier for other people to build, that's good — more engineers running dark factories instead of babysitting cron jobs. JOSH: What's actually running on top of all that right now? ERIK: A hundred and forty-five services on the production box, and a hundred and forty distinct projects have emitted telemetry onto PrimeBus at this point. That's not one app. That's an ecosystem of small things talking to each other. JOSH: Does anything ever step on anything else? ERIK: Constantly, if you let it. That's why every call goes through PrimeRouter instead of agents talking to models directly. One gateway, one place to see the whole picture, one place to pull the plug. [pause] JOSH: Next story — this "Livenerf" thread asking if Opus 5.5 got nerfed. That thing hit 642 points and 256 comments. People are heated. ERIK: People always think a model got quietly weakened after launch. Sometimes it's true, usually it's not — it's routing. Providers load-balance you across different capacity tiers depending on demand, and it feels like the model got dumber when really you just got a different slice of it. JOSH: How would you even know the difference? ERIK: You instrument it. I run PrimeRouter as the gateway for every model call in my fleet — priority tiers, failover, OTEL metrics on every hop. If a model's output quality drops, I can see it in the data instead of guessing from vibes on a forum thread. JOSH: So you've never actually seen a real drop? ERIK: I've seen latency drift and I've seen a backend go quiet and get skipped by failover before I even noticed. That looks identical to "the model got worse" if you're not watching the metrics. Most of Livenerf is that, dressed up as a conspiracy. JOSH: And then there's GPT 6.1 Sol — "near-Astra intelligence for a fifth of the price." ERIK: That's the more interesting story, honestly. Pricing is moving faster than capability right now. When a cheaper model gets close enough to frontier quality, you don't need one model for everything anymore. JOSH: So what does that change for you day to day? ERIK: It's exactly why PrimeRouter exists. Not every call needs the most expensive model. A quick classification task doesn't need the same model as a code review. You route by priority and let cheaper backends handle the low-stakes work, and save the expensive model for the stuff that actually needs it. JOSH: Doesn't that get risky? Cheaper model, worse answer, nobody notices? ERIK: Only if you don't have guardrails. I don't just swap a cheap model in and hope. Anything that touches a protected chain — execution, fixers, review — needs a separate approval before it goes anywhere near production traffic. Cheap and fast is great until it silently breaks something nobody's watching. JOSH: That sounds like a lot of process for swapping a model. ERIK: It's the difference between saving money and creating an incident. Which brings me to the last story. [pause] JOSH: "A Staff Engineer's Guide to Inventing Work." What's that about? ERIK: It's about senior engineers who create process and tooling nobody asked for, just to look busy or feel important. Building for the sake of building. JOSH: That feels like it could describe half of what you do. [beat] ERIK: Fair shot. But there's a difference between inventing work and building guardrails that prevent real work from turning into a disaster later. PrimeBus has run twenty-three hundred ninety-five auto-merge attempts since June 5th. Fifteen-oh-seven got merged automatically. Seven eighty-eight got blocked by Gandalf. JOSH: Seven hundred and eighty-eight blocked? That sounds like a lot of failure. ERIK: That's not failure, that's the guardrail working. Zero of those got escalated to me. Every one of those seven eighty-eight was a change that wasn't safe enough to ship on its own, and the system caught it without waking me up. Sixty-five point seven percent merge rate isn't the story — zero incidents from the blocked ones is the story. JOSH: So the "invented work" is actually the thing keeping you out of the loop. ERIK: Right. The staff engineer trap is building process that just generates more meetings. The version that's worth it is process that removes you from the critical path entirely. I didn't build Gandalf to feel important. I built it so seven hundred bad merges never touch prod and I never have to know about most of them. JOSH: Do you ever go back and check what it caught? ERIK: Every week. If Gandalf's blocking the same kind of thing over and over, that's not the tool being annoying, that's a signal something upstream keeps generating the same mistake. You fix the pattern, not just the individual block. [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: Stop defaulting to your most expensive model for everything. Build a priority tier — critical work goes to your best model, routine work goes to something cheaper, and put a failover in front of both so a rate limit or an outage doesn't take down your whole pipeline. JOSH: How do you actually decide what's "critical" versus "routine"? ERIK: Ask what happens if the answer's wrong. A code review that ships a bug into prod is critical. A classification task that just sorts an inbox is routine. If a mistake costs you money or trust, it goes to the top tier. If a mistake just means you re-run it, send it to the cheap tier and move on. JOSH: And the failover piece? ERIK: That's non-negotiable. Every backend goes down eventually — rate limit, outage, whatever. If your top-tier model isn't reachable, you want the call to quietly drop to the next one in the chain, not throw an error at 2 AM. That's your tip. Use it. [pause] ERIK: If you're building toward financial independence through automation, my first book walks through the whole path. Free chapter at erikandersonbook.com. [pause] 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.