... gate, not just a boolean, which puts it in an awkward middle category your post doesn't quite cover: rules where the check itself needs a model, so you're back to trusting judgment one layer down Comment from Mona Truong(@monatruong_murror): @galdayan  That's the case I skated past, thanks for naming it. A fuzzy rule where the check itself needs a model is a different animal from a boolean guard. The part I'd still hold ...

... news post from 2 October. It's brief, so read it directly instead of relying on my summary. What's the broadest access you've given an agent so far, and would you give it again today? Comment from Mona Truong(@monatruong_murror): @galdayan  Thanks Gal, that's a useful way to put it: the guardrail that held up was on actions, and the read side is the part nobody audits because nothing visibly goes wrong. One cheap thing ... ... reading from acting point is especially important. Giving agents broad access by default can turn a simple prompt-injection issue into a real-world action. Least privilege + human approval for sensitive actions feels like the right approach. Comment from Mona

... evolved. It s the kind of IDE-native documentation layer I always wished existed. I m opening early access for engineers who want to try it and shape the direction. If you re curious, here s the waitlist: codetrackerai.com Mona Hidalgo Founder & Software Engineer Mix It Hub, LLC I d love to hear: What s the biggest code understanding pain you deal with in your IDE? Your feedback will directly influence what we build next ...

... what I d spend the savings on too, I ve seen a bigger context hide the part that actually needs review. Do you keep the small and large runs side by side, or only compare summaries? Comment from Mona Truong(@monatruong_murror): @galdayan  Cost per resolved call without a handoff is a much better yardstick than anything on a pricing page, thank you for putting it that way. The latency spike point is a good one too: averages hide ... ... where the last few entries plus a handful of structured themes did nearly all the work. Do you keep those structured facts as a separate store, or rebuild them from the transcript each turn? Comment from Mona

... absence would stop the product doing one of its functions. If your app is a thin client over your API, that API is not a separate service you can wave at. It's part of the product. This product-versus-service line is the genuinely contested bit, and I'm reading the Commission's guidance and a pile of law firm notes rather than a court ruling. I'm a founder, not a lawyer. If you're near the boundary ... ... ENISA's SRP FAQ. It's the operational one, it's updated regularly, and the last time it changed was 17 September, five days ago. The coverage is not going to tell you about the counter. Comment from Mona

... nobody covered, which is worse, because it changes a line of code a lot of us shipped this year without thinking about it. If your product has a "paste a URL" box, a link unfurler, an enrichment step, a competitor-monitoring job, or anything that fetches a page and hands it to a model, this is about you. What was actually measured. The paper is "Indirect Prompt Injection in the Wild" by Khodayari, Zhang, Acharya and Pellegrino, out of CISPA ... ... probably a floor for voice, not a ceiling. curious if the paper (or anyone here) has looked at whether audio/TTS-mediated delivery changes the compliance rate at all, or if that's just unstudied territory right now Comment from Mona

... right and is more useful than either, and the revised version went up this month, so it's worth twenty minutes before you write your Q4 plan. It's NBER working paper 35275, Demirer, Musolff and Yang - "Writing Code vs. Shipping Code." More than 500,000 GitHub developers, matched against their actual AI usage telemetry rather than a survey, in a matched event study design. The headline numbers are large and I have no interest in arguing with them. Cumulative ... ... they don't all have to be atomic and update every time. I keep them atomic anyway because "catch it up tomorrow" is exactly how AwesomeTTS sat there for months, the second commit never happened Comment from Mona

... Apple's own support pages. Apple In-App Purchase: 26%, or 15% if you're in the Small Business Program, the Mini Apps or Video Partner Programs, or on an auto-renewable subscription after the subscriber's first year. Alternative payment processing inside your app: 20%, or 10% for those same groups. Out-of-app offers with an actionable link: a 15% store services commission, or 10% for those groups, on anything bought within 7 days of the tap. Apps ... ... distributed outside the App Store entirely - an alternative

... Three numbers from it get quoted. Users who saw an AI summary clicked a traditional result in 8% of visits; users who didn't, 15%. Users ended their browsing session 26% of the time with an AI summary present versus 16% without. And the one that should matter most to anyone buying GEO: users clicked a link inside the AI summary itself in 1% of visits. One per cent. That is the click value of the thing the entire optimisation ... ... more often. Just 8% of one- or two-word searches produced one. For searches of ten words or more, 53%. For searches starting with a question word - who, what, when, why - 60%. Sit with that next to the 8%-versus

... times. I'm a founder, not a lawyer, and none of this is legal advice - but the arithmetic of which paragraph applies to you is something you can do yourself in twenty minutes, and it's cheaper than the alternative. Comment from Mona

... could have. And I still don't have a clean answer on the verification-method question, because the honest version is that I don't yet know what I'd do differently with a government-ID-checked 16 versus a self-declared one. I'd rather say that than write a paragraph pretending I've resolved it. The thing I am confident about is narrower. The signal is sparse for a long while yet. Texas only covers accounts created after ... ... refreshing the set from actual refusals in the logs, so users decide what counts as an edge instead of us guessing. Still nothing for the route that gets looser on something no one has tried yet. Comment from Mona

... wrong, mostly they're reasonable, but because a default you never chose becomes a strategy by attrition. This one has a date on it. Comment from Omri Ben-Shoham(@omri_ben_shoham1): Good writeup - the distinction between blocking Training vs blocking Agent is the part people are going to get wrong for a while. One thing worth adding from the operating side of an agent, not the site-owner side - when an agent gets blocked, it usually doesn't fail ... ... built around that, so the failing cases are the ones nobody has a reason to visit. I have no method for generating them rather than stumbling into them. If you have one I would take it. Comment from Mona

... time and the payoff shows up a median of 17 hours later, often more than a day. so the heavy cohort reads as the worst margin right up until you widen the window, and then it flips. Comment from Mona Truong(@monatruong_murror): @siarheihamanovich  The staggered refill is the smartest cheap instrument I've seen for this, because the event is already on the row and it isn't aligned to anything else in the user's life ... ... oddly. The 30-day memory wipe fires per user from their own start date, so I have a natural per-account boundary too. I've never looked at return behaviour around it. Adding that to the list. Comment from Mona

... destination? If the value is a result someone wants and doesn't care how they got an action then being agent-callable is close to free distribution and you should be looking hard at it before your competitor does. If the value is the experience of a person actually being there a destination then racing to become a tool call is optimizing for a future where your product's whole reason for existing has been abstracted away. Worse ... ... ones who shipped a server fastest. They'll be the ones who were honest about whether they should and, if they did, treated that borrowed trust like it was hard to get. Because it is. Comment from Mona

... Mona, a marketer turned founder from Germany. I'm building Xolora, a marketing analytics tool dedicated to solopreneurs and small business owners. I noticed that 80% of that niche is overwhelmed with their marketing and just guessing. When it comes to ads, this means that they're wasting money (around $10,000/year). And the sad truth is that there is no tool out there, that helps them. All the other marketing tools are made for marketers or big teams ... ... underserved niche. Xolora's goal is to simplify ad analytics, give back time, money and ultimately confidence. It just launched in beta and I hope that it will be valuable for its users. What is your niche? Comment from Mona

... loud and instant and mostly noise. In 2026, taking care of the builder is not self-help fluff. It's the actual bottleneck. Treat it like one. Comment from Omri Ben-Shoham(@omri_ben_shoham1): the make mode vs judge mode split is the one that actually changed something for me. i used to check analytics mid build and it wrecked the session every time, not because the numbers were bad but because i'd switch from building to defending ... ... pick it back up next day, or is there some catch-up rule? that's usually where these systems fall apart for me, not the routine itself but the recovery after the first missed day. Comment from Mona

... keeps stuffing it and reliability, not raw capability, is what your users actually feel. Every token you add is a token competing for attention. Spend them like they're scarce, because to the model, they are. Comment from Mona Truong(@monatruong_murror): @omri_ben_shoham1   Yes that's the part I underplayed in the post. Curating once is a prompt; re-deciding every turn is a system. What's helped us is treating context assembly as a step that ...

... this: if building is no longer the moat, then the moat is available to anyone willing to do the unglamorous work. Knowing your person. Owning a relationship instead of renting attention. Being early to a workflow before a competitor gets embedded in it. None of that requires funding or a team. It requires you to stop hiding in the part of the work that feels productive and go do the part that actually decides whether the thing lives. Ship ... ... abstract: the AI coding tools that spread fastest did it through something people could point at a generated PR, a public run not an invite-a-friend prompt bolted on after launch. Probably the cleanest version of "output vs

... best day, on a task you made up. Your users meet it on an average day, on the task they actually have. Build for the average day. That's the whole game now. Comment from Mona Truong(@monatruong_murror): @dalemooney   This is a sharper version of my third point than I wrote, and you're right that it has the same shape as the problem a report sensor only fires when the user already caught ... ... transcripts starts accurate and then decays, not because the model changed but because your users did. Which means it has to be refreshed from live traffic on a schedule, or in a year you have rebuilt the exact lab-versus

... inside your app if they can only feel understood while they're talking to your model you haven't solved loneliness. You've made it more comfortable to stay lonely. You've built a very sophisticated waiting room. The alternative is to design for the exit. Concretely, that means three things. Point users back at real people, not deeper into the app. The most valuable thing an emotional AI can do isn't be a better listener than your friends ... ... about: when your user closes the app, are they better off, or just less alone for the next four minutes? Your roadmap already knows the answer. The question is whether your metrics will let you admit it. Comment from Mona

... which one were you actually optimizing for? Comment from Nika(@busmark_w_nika): imo, the whole thing with AI agents is becoming oversaturated the same way as it was with AI in 2024 Comment from Stacy Wycoff(@stacywycof83995): Mona, this is the question I keep circling back to while building FounderFlow. My honest answer is uncomfortable, I want usage to go down over time, not up. If it is doing its job, a founder should need fewer of my alerts ... ... like proof of value to an investor, but for the actual founder using it, quiet is the win. How are you thinking about measuring a product that is supposed to make itself less necessary? Comment from Asad M.(@asadmalik901): Mona

... refuse to build, even when users are asking for it and the metrics say yes. Anyone else held the line on a feature like this? Curious how you explain the "no" to users who genuinely want it. Comment from Mona Truong(@monatruong_murror): @asadmalik901  "Reading it as a report of a moment" that reframe is going to stick with me. You're right that it changes everything about how we respond. We've actually started experimenting with exactly what ... ... framing is spot on. We've been saying what we don't do instead of what we protect. "The AI waits until you're done" is a much better way to explain it. Thank you for this. Comment from Mona

... animation and 8 seconds of silence. Our engineer thought it was a bug when he first saw the spec. "You want me to add... a delay?" Yes. On purpose. Here's why. We ran an experiment: instant AI response vs. an 8-second pause before the same response. The content was identical. The only difference was time. Users who got the pause rated the AI's reflections 40% more helpful. Same words. Same insights. They just felt more considered ... ... before showing an AI response makes it feel more thoughtful and natural. It's a good reminder that a better user experience isn't always about making AI faster, it s about making it feel more helpful. Comment from Mona

... language explainer of a chunk of the codebase, not because you need it day to day, but so there's a paper trail you can hand your CTO when something breaks and speed matters more than pride Comment from Mona Truong(@monatruong_murror): @omri_ben_shoham1  that 2am scenario is exactly what keeps me up honestly. you're right that the middle ground matters I've been getting better at asking the agent diagnostic questions like "what could cause ... ... data flow here." it's not the same as reading the code yourself, but it's a different kind of literacy. learning to debug through conversation instead of through code. still early in figuring that out though. Comment from Mona

... health aren't always the same direction. Sometimes the right metric is one that goes down. Anyone else building products where you've had to choose between engagement and user wellbeing? Curious how others navigate this. Comment from Mona Truong(@monatruong_murror): @galdayan  honestly it was uncomfortable. our first reaction as a team was denial we kept looking for a bug in the analytics. when we confirmed the drop was real, there was a moment where it felt like ... ... small team where everyone believed in the mission enough to stomach a dip. I think if we'd had external pressure from investors optimizing for growth metrics, it would've been a much harder conversation. Comment from Mona