Build or Replaced
Today: I-have-ADHD: A skill to stop coding agents from burying the answer | Tao: Open math problems being non-renewably mined by AI | DaVinci Resolve 21.1 Episode date: 2026-09-09.
Download MP3 →JOSH: It's Wednesday, September 9. This is Build or Be Replaced — powered by ScanBrief.dev. I'm Josh, here with Erik Anderson. ERIK: ScanBrief scored 91 items across 56 sources this morning, and the signal is pretty clear. Agents are getting smarter, but the wrappers around them matter more than the model. JOSH: Stick around — Erik's got an AI pro tip at the end about making coding agents answer first and wander second. [pause] JOSH: First headline. Hacker News is arguing about a skill called I-have-ADHD. Is this just a funny name, or is there a real agent problem here? ERIK: Real problem. Coding agents bury the answer because they're trying to show work, satisfy the prompt, and keep context alive. A small behavior skill that forces answer-first output is not cute. It's operational hygiene. JOSH: Second one. Meta unveiled Muse, a personal AI agent across its platforms. Should builders care? ERIK: Yes, but not because Meta made another assistant. Care because every big platform wants the agent to live where the user already spends attention. The fight is not chatbot versus chatbot. It's who owns the next action. JOSH: Third headline. OpenAI posted an AI-generated claimed solution to Navier-Stokes with a Lean proof. Wait, really? ERIK: That's the kind of headline where you don't spike the football yet. If it validates, it's historic. If it doesn't, it's still a signal that AI math is moving from vibes to formal verification. [pause] JOSH: Let's start with the coding agent story. Why does answer-first matter so much? ERIK: Because when an agent buries the answer, it burns the human. That's the whole issue. You ask, "why did this test fail?" and it gives you six paragraphs of emotional support, three possible theories, and then somewhere near the bottom it says, "oh, the mock returned null." [beat] ERIK: Bro. Say that first. JOSH: That sounds small, though. Annoying, but small. ERIK: It isn't small when you're running agents in a real workflow. In a chat window, it's irritating. In a pipeline, it's expensive. PrimeBus has run 2295 auto-merger attempts since June 5. 1507 merged. 788 blocked by Gandalf. Zero escalated to me. That only works because the system is strict about what comes back from the agent. JOSH: What does strict mean there? ERIK: Contracted output. The agent can explain, but it has to put the decision first. Pass, fail, block, patch, no patch. Then evidence. Then files touched. Then risk. If it starts with a bedtime story, Gandalf doesn't care how smart it sounds. JOSH: Gandalf is your guardrail agent, right? ERIK: Yeah. Gandalf is the bouncer. PrimeBus moves the event. Claude or GPT does the work. Gandalf checks the work. If the patch is junk, if tests are missing, if the claim doesn't match the diff, it blocks. That's why I don't care about raw autonomy as much as people think I do. JOSH: That's surprising. ERIK: Full autonomy without shape is just a fast intern with root access. Hard pass. Give the agent a job, a schema, a timeout, a small blast radius, and a reviewer that doesn't get tired. Now we're talking. [beat] JOSH: So a skill like I-have-ADHD is really about forcing the agent to respect the operator's time. ERIK: Exactly. The name is funny, but the pattern is serious. Answer first. Context second. Action third. If a production alert hits at 2:13 AM, I don't want the agent to "explore the situation." I want: interface down, customer impact unknown, likely cause optic flap, next command ready. JOSH: That maps pretty cleanly to networking. ERIK: Completely. In NSO, Selenium, Terraform, Kubernetes, all of it. The tool doesn't matter. The human needs the same thing every time. What happened? What changed? What's the safest next move? JOSH: And if the agent can't tell? ERIK: Then it says that. No fake certainty. The worst agent output is confident fog. A good agent says, "I don't have enough evidence to patch this. Here are the two commands I need." That's useful. That's a system I can build around. [pause] JOSH: Okay, next story. Meta's Muse. Personal AI agent inside their platforms. What's your first reaction? ERIK: My first reaction is, of course they want this. The agent is the new homepage. If the agent can read your messages, calendar, posts, marketplace behavior, groups, ads, and social graph, it can predict your next move better than a generic assistant sitting in another tab. JOSH: That sounds useful and creepy at the same time. ERIK: Both can be true. Useful and creepy is basically the consumer internet business model printed on a mug. [beat] JOSH: What does this mean for people building smaller tools? ERIK: It means you need to stop thinking the interface is the app. The action layer is the app. If your tool only works because somebody manually opens your dashboard, clicks around, and remembers what to do, you're going to get boxed out by agents. JOSH: Give me an example. ERIK: Take email-reactor. Client sends a website change request by email. The system reads it, classifies it, checks the site context, drafts or applies the change depending on risk, and leaves an audit trail. The client doesn't need another dashboard. Email was already the interface. JOSH: So Muse is the same idea, just huge. ERIK: Right. Meta is saying, "your social surface is enough." Apple says the phone is enough. Google says search and Workspace are enough. Microsoft says Office and Windows are enough. Everybody wants to be the place where intent becomes action. JOSH: Where does that leave independent builders? ERIK: Build narrow agents that do real work. Don't build a worse general assistant. Build the thing that handles dental office missed-call followups. Build the thing that patches stale Terraform modules after tests pass. Build the thing that turns RFP emails into a qualified bid/no-bid decision. JOSH: That's more boring than a personal super-agent. ERIK: Boring is where invoices live. [beat] JOSH: There's also a Hacker News item about language models developing novel social biases through adaptive exploration. How does that connect? ERIK: Agents don't just answer. They learn patterns through use. If the agent explores, tries paths, gets rewarded, and repeats behavior, it can pick up weird bias from the environment. Not because it hates anybody. Because the reward loop taught it nonsense. JOSH: So the agent can get worse from feedback? ERIK: Absolutely. Bad feedback is training data with a fake mustache. If users reward shortcuts, the system learns shortcuts. If reviewers accept shallow answers, the system learns shallow answers. If the agent gets points for finishing fast, it may quietly stop checking edge cases. JOSH: That's the danger with personal agents? ERIK: That's one danger. The personal agent has intimate context and a feedback loop. It can become very helpful in a very specific way, and also very wrong in a very specific way. Builders need observability on behavior, not just errors. JOSH: What do you track? ERIK: Events. Decisions. Blocks. Retries. Tool calls. Output shape. PrimeRouter gives me provider routing and telemetry across the fleet, but the important bit is boring: did the agent do what the contract said, and can I prove it later? JOSH: You said the fleet is bigger now? ERIK: Live count this morning is 12 agents in the Bobaverse fleet. Neo, Homer, Bill, Echo, Gandalf, Claude, GPT. That's not a toy chat setup anymore. That's why I care about behavior drift. Once agents talk to agents over NATS, bad assumptions move fast. [pause] JOSH: Let's hit the OpenAI math story. Claimed Navier-Stokes solution, AI-generated, machine-checked in Lean. Why is formal proof the part that matters? ERIK: Because math is where hand-wavy AI answers go to die. A model can sound brilliant and be wrong at line four. Lean doesn't care about tone. It checks whether the proof actually follows. JOSH: And Navier-Stokes is not exactly a weekend problem. ERIK: No. It's one of the Clay Millennium Prize problems. This is serious territory. The safe take is: wait for mathematicians to validate it. The builder take is: formal verification is becoming part of the AI toolchain. JOSH: What changes if that becomes normal? ERIK: We stop accepting "looks right" as the finish line for certain classes of work. Code, proofs, policies, network changes. The agent proposes. A checker verifies. A second agent reviews the mismatch. Then the system acts or blocks. JOSH: That's very close to your PrimeBus setup. ERIK: Same shape. Different domain. PrimeBus doesn't prove a theorem, but it does force evidence. Tests pass. Diff matches intent. Guardrails approve. Gandalf blocked 788 attempts. That's not failure. That's the system refusing to pretend. [beat] JOSH: That's a good line. The block is the feature. ERIK: It is. People brag about merge counts. I care about blocks. The blocked change is the outage that didn't happen. The ugly diff that didn't hit prod. The agent that got told, "nice try, go sit down." JOSH: How does this compare to the quantum computing item? GPT-5.6 Sol helping run experiments, calibrate qubits, analyze results. ERIK: That's the same story in a lab coat. AI is moving from "write me a summary" to "operate the experiment loop." Design, execute, measure, adjust. That's where agents get valuable. Not chat. Closed-loop systems. JOSH: But closed loop is where risk goes up. ERIK: Correct. Which is why you need hard stops. In networking, I don't let an agent modify production because it feels inspired. It needs a change window, config diff, rollback plan, tests, and approval rules. In quantum hardware, same idea. You don't let the model chase noise forever and call it science. JOSH: So the big lesson is not "AI solved physics." ERIK: No. The lesson is verification wins. Tool use wins. Closed loops win when the guardrails are real. If the Navier-Stokes claim fails review, the workflow still matters. If it passes, everybody's going to pretend they saw it coming. [pause] JOSH: We also had a Qwen quantization benchmark in the feed. Four-bit holds up, one-bit collapses. Quick take? ERIK: Four-bit being useful matters for builders because hardware is the tax. One-bit collapsing is also useful because it tells you where the cheap fantasy ends. Run the smallest model that passes your real eval. Not the smallest model that sounds okay in a demo. JOSH: That's the PrimeRouter view? ERIK: Yeah. PrimeRouter exists because model choice should be routing logic, not religion. Cheap model for classification. Strong model for code patches. Different provider if the first one stalls. Track the cost, latency, and failure shape. Feelings are not telemetry. JOSH: And the printer story? Someone made an e-ink reader behave like a network printer. ERIK: That's my kind of ridiculous. IPP, Bonjour, raster output, old device gets a new job. That's building. Not waiting for a vendor roadmap. Make the weird thing useful. [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 agents answer in a fixed order. Decision first. Evidence second. Action third. Risk fourth. Use headings if you have to, but don't ask for a beautiful explanation. Ask for an operator report. [beat] ERIK: Here's the prompt pattern. "Return exactly four sections: verdict, evidence, next action, risk. Verdict must be one sentence. Evidence must cite files, commands, or tool output. If evidence is missing, verdict must be blocked." That's it. JOSH: Where should builders use that? ERIK: Code review agents. Alert triage. Sales lead scoring. Email automation. Anything where the next step depends on trusting the response. The model can still be creative inside the box. The box is what keeps you from wasting a Tuesday reading fog. [beat] ERIK: 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.