... 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 ... ... does not work. The ceiling is 8%, and that is small open models on the one representation that hides the trick. Closed-source and medium models sit at 0.6%. Meanwhile the string is permanently in your markup, it is trivially

... against a half-drawn screen. Too long and every test crawls. So we don't sleep. We watch for the screen to stop changing and move when it has. There's no sleep(2) anywhere in the executor. Sounds trivial written down. It's most of the difference between tests that pass and tests that pass every time . 4. Knowing when to stop An agent with a retry loop and no budget will burn twenty minutes ...

... still routes through one person. This week we shipped something I'd been putting off for months, because I suspected the answer would be embarrassing. We had no way to know whether our recommendations helped anyone. Generating advice is trivial. Proving that someone saw it, acted on it, corrected it, or got a result is hard, and we'd been shipping the easy half. Three things went out. A permanent record of every meaningful recommendation, so it doesn't vanish ...

... asking. Then the migration is changing what that lookup returns, in one place, rather than auditing every screen that said "your projects" The part nobody warns you about is the rows you already wrote. Backfilling an org id is trivial. Deciding which org owns a row created before organizations existed is what eats the week, and the fun case is two individual accounts that turn out to be one customer, with data on both sides that somebody merges by hand ...

... solid answer to that. One scenario I don't think you covered: what happens when the underlying source document itself gets silently edited rather than replaced? Like someone fixes a typo in a pricing page and the diff looks trivial to a human skimming it, but the agent's approved answer was actually anchored to the old wording in a way that changes the meaning. Drift detection based on the agent's own output might miss that because the agent ...

... regardless of amount, because a new payee is the actual risk, not the number. Same logic on messaging, an internal note runs free, anything that goes out publicly in my name gets a gate even if the content is trivial. So the verb decides which axis matters, then the specific conditions on that axis decide autonomous versus

... articles a day for my service, and traffic to the site has grown around eight times in the last six months. The waste, then, is the obvious trap inside that number. Volume for its own sake. It is now trivial to have AI produce a thousand articles nobody needed, and that buys you nothing except a slower site and a worse reputation with the search engine. Most of my work is not generating articles, it is the system that decides ...

... version is. Nothing reaches training without passing a blocking check. If a failure was detectable before we charged you, it isn't allowed to happen after. Every error comes back with a cause and something you can click. Sounds trivial written down. Was most of the work. What it doesn't do: anything other than text, pick between architectures for you, or improve the model after deployment. All on the list. I'd rather say it than let the landing ...

... That slot machine feeling is so real when users burn credits on bad outputs. I've seen a few teams use a free retry token for failed jobs instead of touching credit balances. How are you detecting system failures vs bad prompts on your end? Comment from Asad M.(@asadmalik901): @galdayan Fair, and abandoned one step later is a real hole. What makes me okay with it for now is that the drag is deliberate and the abandon ... ... walk-aways is a much cheaper mistake than overcharging someone who actually got value, and it's the kind of thing you can tighten later once you have real usage data on how often people download-and-leave versus

... solution instead of the quickest one. Review: finding related edge cases and preventing the issue from recurring elsewhere. Unplanned work: handling interruptions and problems that were never part of the original task. When this context disappears, complex work looks trivial and efficient problem-solving becomes almost invisible. Teammates see what changed, but not why it changed, what alternatives

... small change has made my AI coding sessions much easier to control. Before asking the agent to implement a non-trivial feature, I now ask it to write a short implementation plan first. Not a huge specification. Just: What files it expects to change What existing code it plans to reuse What assumptions it's making What could break How it plans to verify the change Then I review the plan before giving it permission to implement anything. I started ...

... mobile carrier that's a lot of separate people behind one address, and the first one through the door locks out the rest. Anyone who actually wants around it changes network and carries on. Cookie and localStorage counters. Cleared trivially, so they only stop the people who weren't trying. Not worthless, but a speed bump rather than a limit. Browser fingerprinting. Would probably work. I don't want to do it. It feels like a worse thing ...

... finally read the whole study instead of the headline, and the actually useful findings are different from what got shared. First, in their realistic-config lane (instruction file + several MCP servers), the "light" harness came out HEAVIER: ~90.8k vs ~75k. A 72KB instruction file alone added ~20k tokens to every request, on both harnesses. Their own conclusion: configuration, not the harness, accounts for most of the production bill. The harness sets the floor, you set the ceiling. Second ... ... close to a fixed cost. An unstable small one can out-spend it. Size isn't the sin, churn is. Third, session shape flips the winner anyway. On a multi-step task the heavy harness finished cheaper (121k vs

... test suite. The first install. Five things broke on our first paid setup. Three I can describe cleanly: The customer refused to hand over the domain. Our whole provisioning model assumed a transfer, because that made certificates and routing trivial for us. He had email records on that name and had known us for about a week, and he was right. We rewrote it to apex-only, so the customer keeps their own DNS and we never hold the name ...

... codebase less than I used to - I'll touch one file and have no real sense of what else depends on it until CI (or prod) tells me. Ended up manually tracing imports before merging anything non-trivial, just to make sure I'm not breaking something three files away. Anyone else running into this, especially with AI-assisted code? Curious if it's a real pattern or just me being paranoid. Comment from Hossein Yazdi(@hosseinyazdi): Yeah, I think ... ... from Alper Tayfur(@alpertayfurr): I don t think you re being paranoid. When AI writes most of the feature, we skip some of the slow work that normally builds our mental map of the codebase. Before merging anything non-trivial

... anyone else tracks this specifically, or if there is an earlier signal worth watching for before the third person even shows up. Comment from Abdullah Javaid(@abdullah_javaid3): @jernej_jan_kocica  Haven't coded each reply as agreed versus pushed back specifically, so this is memory, not an audit. The 9 view thread was mostly me agreeing with whatever came in. The 221 one had at least one spot where I admitted not controlling for a variable someone raised ... ... javaid3): @alieksia  Checked before answering. The 221 thread's actual question was "do you have a way to predict this, or is it trial and error too," pure opinion, nothing to look up. It still compounded, so fact versus

... here's what we do, here's where we're unreliable, here's the thing we put in front of you before it goes out. That's a harder page to write and a much harder one for a competitor to copy, because it means you actually went and found your failure modes. Curious whether anyone has tried this and what it did to conversion. My guess is it costs you signups and buys you retention. Comment from Asad ... ... already have the number and have never looked at it because nobody thought to ask. Comment from Asad M.(@asadmalik901): @alpertayfurr  The reason almost nobody shows it is that they never built it. Undo on a draft is trivial

... Context Window Compounding: In long-running agentic sessions, the entire conversation history and tool output (often 100,000+ tokens) are re-transmitted on every single iteration. As a result, token consumption scales exponentially, not linearly. 2. Delayed Cloud Governance vs. Real-Time Execution: Native provider dashboards (OpenAI, Anthropic) primarily offer delayed usage alerts rather than zero-latency pre-request validation. In an automated test loop or infinite tool retry cycle, an agent can burn through hundreds of dollars before ... ... cloud threshold email is dispatched. 3. Asymmetric Cost Structures (Prompt Caching): Modern architectures like Anthropic s prompt caching introduce non-trivial

... understood it, mostly by necessity. Now understanding is a separate step you have to choose to take, not something that comes free with writing the code yourself. One habit that closes that gap fast: before merging anything non-trivial, ask the agent to explain its own implementation back to you, line by line, in plain language. Not as a formality, actually read the explanation. Two things tend to happen. Either it confirms your understanding and takes thirty seconds ...

... nothing, the reviewer's normal experience becomes an empty screen, and empty screens train people to expect empty. Whether the one that is not empty gets read properly on day forty is an open question. It still beats the alternative, because a difference against a background of silence is about as salient as anything gets. And yes to the small pile. Most of what gets called judgement is work nobody has bothered to make mechanical yet. Comment from Omri ... ... Uddin(@hammad_shams_uddin): The gate on reach is right, but I think there is a third case it still misses and it is the one that got me. Some of the worst reach comes from actions that are trivially

... first group is the one that needed the fingerprint, and it's the group i'd bet on long term: it survives renames and re-obfuscation for free Comment from Giovambattista Fazioli(@gfazioli): @bill_frad  thanks "informative vs genuinely useful" is a better framing than the one i had, might steal it. the thing that tipped it for me was actually the boring case: if there's no baseline yet, the report says so out loud. sounds trivial

... worker. For UI testing, a remote headed browser is not just a convenience. It changes the debugging loop. Instead of asking "what did the script think happened?", you can ask "what did the user actually see?" ### Harness Code vs Local Agent Harness Another important decision is where the agent capability should come from. One option is to build the entire reasoning layer yourself: call an LLM API directly, manage prompts, retries, tool calls, screenshots, JSON parsing, and safety rules. That ... ... gives maximum control, but it also means the test harness becomes an AI platform project. The alternative

... already scores every stuck URL by value, so with an API it would submit the top ten each morning and then check a week later which ones actually got indexed, closing the loop without me. The click itself is trivial. The part worth automating is never forgetting to do it. Comment from Brady Wilfahrt(@bradywilfaqn): why do you think Google hasn't released this API yet? Comment from Abdullah Javaid(@abdullah_javaid3): @tony_tan10086  Really like how you distinguish ... ... input that makes the ranking self-correcting. If a money page sits in Discovered for three weeks despite a manual request, that's not a queue problem you can click your way out of - it's the crawl-budget-vs

... than that. Comment from Jernej Jan Ko ica(@jernej_jan_kocica): Different domain, same problem, and it breaks the taxonomy a bit. My agents don't write code, they talk to customers. Every list in this thread sorts by reversible vs destructive. Deploys, db writes, rm. By that sorting, "send a reply to a customer" is about as safe as it gets. It's text. It's not a deploy, nothing gets erased, no invariant breaks. It's also ... ... whole system I cannot undo. There is no rollback on a message someone already read. Wrong refund promise, wrong delivery date, and the damage is a person, not a branch. So for us the line isn't destructive vs

... constraint I had repeated twice. After that, I started checking every done myself. Comment from Ghost Kitty(@ghost-kitty-9972846): Comment Deleted Comment from Neeraj H(@terranicgerald): @amberburch399  Yeah, that's exactly it, the rename itself was trivial to fix. What actually cost me was that it reported success with the same confidence whether it had checked one file or all thirteen. There was no signal to tell those two situations apart. I like the idea of surfacing uncertainty ...