--- title: "Issue 1" canonical: "https://current.tinyfish.ai/issue/1" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/1" issue: 1 description: "Systems flip when nobody's watching—agents take the wheel, teams hoard double what they need, and reasoning models reason in circles. This week: the inversions reshaping how work actually works." authors: ["Sable Whitford","Nora Kaplan","Rina Takahashi","Sevda Polat"] published: 2025-10-28 updated: 2026-11-05 articles: 77 --- # TinyFish > CURRENT is a journal about AI agents in production, for the engineering and product leaders who have to make them work. Drawing on experience building enterprise web agents, it explains why the demo succeeded and the deployment didn't, and what that means for choosing vendors, briefing the board, and avoiding mistakes others have already made. [Read on the website]() ## Issue 1 Systems flip when nobody's watching—agents take the wheel, teams hoard double what they need, and reasoning models reason in circles. This week: the inversions reshaping how work actually works. ### [Daily Brief]() Your daily digest of web developments through the lens of enterprise agent deployment - [Google Lets Schools Keep Search Without Gemini. The AI Overviews Come Along Anyway.]() — By [Sable Whitford]() · Published 2026-11-04 On Tuesday, which was also Election Day, Google announced a K-12 policy that switches Gemini, Gemini Notebook and **Search off by default** for schools using its education software. A Google education VP told the New York Times that "Search and AI are deeply integrated," so districts that ban AI would lose Search as well. Bernie Sanders called it **"extortion,"** and Sen. Richard Blumenthal and NYC Mayor Zohran Mamdani joined in. Within hours Google **partly walked it back**: schools can now turn on Search by itself, without Gemini. That Search still puts AI Overviews on the results page, though, so **the fight now is over what "AI off" means.** - **Proxy Traffic Leaving the White House Network?** — [Source]() — Published 2026-11-04 Layer3 Intel says residential proxy traffic exited AS2650, the Executive Office of the President's network, 236 times in 30 days through 28 proxy services. It says it wasn't even looking. Its guess is that someone working late has a device carrying a proxy SDK. Treat that as a theory: no device has been named, and six public IPs can sit in front of a lot of machines. - **DreamHost Had Uninvited Guests Since August** — [Source]() — Published 2026-11-04 DreamHost confirmed that attackers used a file manager bug to steal session tokens, customer emails and files, and that they got in back in August. That's a long stay at a host with about 400,000 paying customers. DreamHost says it's unrelated to PHP2Shell. If you're a customer, rotating credentials today is a sensible move. - **PHP2Shell's Hangover Gets a Sequel** — [Source]() — Published 2026-11-04 PHP2Shell was a pair of 9.2-severity bugs that put NASA, Wikipedia and the White House at risk. Two weeks after the patches, Kaspersky counts nearly 30 million servers still unpatched, plus a third, still-open flaw (CVE-2026-115560). Working exploits keep multiplying, though there's no confirmed mass-breach wave yet. Tolmo's Matt Suiche says AI bug hunting is exposing 40 years of legacy code, and that's today's big argument. AI can find a bug in minutes. Someone still has to book maintenance windows for 30 million servers. - **Can DNS Tell You It's a Bot?** — [Source]() — Published 2026-11-04 Singapore's IMDA is reportedly exploring, with GoDaddy, Cloudflare and Google, whether DNS could identify AI agents. That would give the internet's 1983 phone book a second job checking IDs at the door. Every ops person already knows where the postmortem will land: it's always DNS. - **Meta Buys Conductor, Cloud Users Get Notice** — [Source]() — Published 2026-11-04 Meta bought Conductor, the app for running several coding agents in parallel, and the team is joining Meta Superintelligence Labs. Terms weren't disclosed. Conductor Cloud goes read-only Nov. 13 and shuts off Nov. 20. The local app gets model updates and critical fixes until early 2027. Conductor ran agents from rival labs too, so everyone wants to know whether that survives the move. - **OpenAI Shops Too: Pierre Computer** — [Source]() — Published 2026-11-04 Not to be outdone, OpenAI told Axios it plans to buy Pierre Computer, which stores and manages AI-generated code. It's OpenAI's second coding pickup after the Ona deal announced in June. Agents are producing mountains of code nobody typed by hand, and someone has to keep track of it. OpenAI would like to be that someone. - **Kimi K3.1 Pro Weights Drop: Bring Disks** — [Source]() — Published 2026-11-04 Moonshot launched Kimi K3.1 Pro in October, and today the open weights landed on Hugging Face. It has 2.8 trillion parameters with 104B active, and reportedly ships as 96 shards totaling about 1.56 TB under a custom non-OSI licence. Self-hosting starts with a purchase order for storage. Before the release, Artificial Analysis said it would cut the gap between open and proprietary models from 12 points to 3. - **Ripple-65B Crowns Itself America's Coding Champ** — [Source]() — Published 2026-11-04 Ricursive calls its Ripple-65B Preview "U.S. Open-Weight SOTA on Agentic Coding," meaning best in its class. That's the company grading its own homework. The "U.S." qualifier is carrying extra weight today, since a Chinese lab just released 1.5 terabytes of open weights. - **Go Gets a Brand-New JSON API** — [Source]() — Published 2026-11-04 The Go team introduced encoding/json v2, a new official JSON API, and Gophers have plenty to say about it. Nearly every Go service reads or writes JSON somewhere, so a second version of the standard package lands on the desk of anyone shipping APIs. - **Apple Puts Engineers on an AI Diet** — [Source]() — Published 2026-11-04 Bloomberg reports that in October, Apple cut some engineers' daily AI spend caps from $300 to $200. Some could drop to $100, and certain non-engineering teams could be capped at $25. Some teams are being nudged toward cheaper or on-device models. The punchline: few people ever hit $300, and Apple's R&D rose 32% in fiscal 2026. - **Texas Voted, and Data Centers Flinched** — [Source]() — Published 2026-11-04 Last brief, the data center backlash was winning local ballot fights. Now it has the Texas governor's mansion, in a state holding 25%+ of planned US capacity. Gina Hinojosa ran on an immediate moratorium, tougher water and power rules, and no tax breaks, and AI data center stocks fell on the result. GLJ Research says about 13 GW is exposed. Altimeter's Brad Gerstner says 10–15 GW might never get built. Abbott had already paused permits. Musk, who spent big on the losing side, now likes "data centers in space." - **Big Tech Pulls Out the AI Credit Card** — [Source]() — Published 2026-11-04 Oracle filed to sell $20B of new stock and bonds for its AI buildout, and Meta is seeking about $25B in bonds. The timing is bold. They're raising money to build on the same day the state with the most planned capacity elected a governor who ran on a moratorium. - **ChatGPT Search Is 20% of Google's?** — [Source]() — Published 2026-11-04 OpenAI's Adam Fry said Monday that ChatGPT search volume is about 20% of Google's, and that weekly users reportedly went from 1B to 1.2B in 60 days. Check the counting rules first. Any prompt that triggers a web search counts, and OpenAI didn't disclose the Google baseline or its method. Ahrefs puts ChatGPT at 0.36% of web referral traffic. - [Lebanon's Two Clocks and the Question Calendars Don't Ask]() — Published 2026-10-30 On March 26, 2023, Lebanon woke up on two clocks. The caretaker government had postponed daylight saving three days earlier. Churches and TV channels moved forward anyway, government offices stayed put, and most phones jumped on their own. Middle East Airlines printed every departure an hour earlier, so its planes still left at the same moment. That week made each calendar show what it had been storing: an exact instant, or a time on the wall that politics can move. Software assistants now make that choice without asking anyone. - [Your Agent's Questions Need a Recipient and a Timer]() — Published 2026-10-30 When an AI agent stops mid-task to ask a human something, it often reads as a failure of autonomy. In one study, letting coding agents ask about bug reports stripped of key details won back most of the lost ground, lifting solve rates from about 55% to 69.4%. But the "human" answering was another model, holding the full report and standing by. Real people are in meetings, or asleep. The design work is in the ask: who gets it, what it says, and what happens while it waits. - [Faria Huq and the Reason Behind a Human Step]() — Published 2026-10-30 When a person takes control from an AI agent driving their web browser, the system records a human step, but not why. Faria Huq and her Carnegie Mellon collaborators have built tools around these moments: an agent that can't find the right forum, a shopper who realized the brand mattered only after seeing the agent's pick. Their system learns to predict when someone will step in. The reason it sets aside would tell a team whether to fix the agent, the website, or its understanding of the person. - [Who Pays While the Agent Waits]() — Published 2026-10-30 Google's new service for hosting AI agents starts from a fact anyone who has watched one run already knew: most of the time, the agent is waiting, for the model to answer or for a person to reply. So Google suspends idle agents and promises you won't pay for idle machines. Fair enough, but the wait still gets paid for. It moves into approval queues nobody times, and then onto the plan, which comes back exactly as it left, even if the stock it counted on has since sold. ### [Market Pulse]() Reading the agent ecosystem through a practitioner's lens - [The Eight-Hour Agent Nobody's Running]() — By [Nora Kaplan]() · Published 2025-10-27 · Format: The Signal Been tracking AWS's eight-hour runtime thing since July—longest in the industry, apparently—and the weirdest part is nobody's talking about actually using it. Spent enough time thinking about collaboration tools to know: trusting software to work unsupervised for a full business day requires organizational muscles most teams haven't built yet. The infrastructure's here. The confidence? Nowhere close. Feels like watching storm clouds roll in while everyone's still deciding if they need an umbrella. Capability exists, readiness doesn't. - [The Velocity Mismatch]() — By [Nora Kaplan]() · Published 2025-10-27 · Format: Field Notes Been watching this pattern for weeks now and it's getting clearer—teams are deploying agents way faster than they're building the governance infrastructure to manage them. McKinsey dropped numbers this month showing 80% already running agents, 96% planning expansion, but when you ask about barriers it's all governance tools. Classic mismatch between what moves fast (spinning up agents) and what moves slow (audit trails, access controls, observability across systems). - [When Infrastructure Becomes a Category]() — By [Nora Kaplan]() · Published 2025-10-27 · Format: Dual Lens Kernel raised $22M for browser automation infrastructure this month. Not remarkable because it's novel—remarkable because it's becoming its own category. When Perplexity pays for browser automation instead of building it, that's the signal. Something crossed from "feature we should own" to "infrastructure someone else handles." Spent years watching this pattern: observability, auth, now browser automation. The complexity isn't in spinning up Puppeteer. It's in making it work reliably when websites actively resist you. First piece tracks when that threshold gets crossed. - [The Problem Infrastructure Can't Solve]() — By [Nora Kaplan]() · Published 2025-10-27 · Format: Dual Lens Second piece is the uncomfortable part. All that infrastructure—session management, CAPTCHA solving, fingerprint rotation—exists because we're making agents pretend to be humans. The web resists programmatic access by design. Not a bug. An architectural reality accumulated through millions of decisions about bot prevention, personalization, security. Infrastructure can wrap that complexity. Can't resolve it. Can't make websites behave like APIs when they're fundamentally experiences testing whether you have eyes and a mouse. We're not fixing the web. We're learning to work around what it became. - [AgentKit Promises Complete Tools, Delivers Partial Assembly]() — Published 2025-10-27 OpenAI launched AgentKit as "a complete set of tools for developers and enterprises to build, deploy, and optimize agents." The pitch: solve fragmented tooling, eliminate weeks of frontend work, cut development time by half. - **OpenTelemetry Conventions for Agent Observability** — [Source]() — Published 2025-10-27 OpenTelemetry is establishing semantic conventions for agent telemetry. The payoff: compare performance across frameworks, switch implementations without rebuilding your observability stack. Interoperability at the infrastructure layer means your monitoring investment survives framework churn. - **Microsoft Merges AutoGen and Semantic Kernel** — [Source]() — Published 2025-10-27 The Agent Framework public preview consolidates research tooling with enterprise foundations. Built-in observability and durability as native features, not afterthoughts. Fewer frameworks to evaluate, more reliability assumptions you can count on. Maturity looks like convergence. - **Evaluations Join Metrics as Core Observability** — [Source]() — Published 2025-10-27 Azure now treats evaluations and governance enforcement as observability capabilities, not separate concerns. Non-deterministic systems need systematic quality assessment to run in production. Traditional metrics tell you what happened. Evaluations tell you if it was correct. - **Governance Tooling Blocks AI Adoption** — [Source]() — Published 2025-10-27 McKinsey found governance tooling is the top barrier to AI adoption. Only 26% of organizations have integrated GenAI standards into governance frameworks. Task adherence, prompt shields, PII detection determine what ships. Demos are easy. Compliance is the gate. - **EU AI Act Enforcement Underway** — [Source]() — Published 2025-10-27 The EU AI Act started enforcement in August 2024, full compliance expected by 2026. Risk-based classification means high-risk applications face stricter scrutiny. Architectural decisions about data governance, bias prevention, and audit trails are happening now, not later. - **SRE Agents Built Into Observability Platforms** — [Source]() — Published 2025-10-27 Grafana's Assistant investigations and Middleware's OpsAI shift monitoring from reactive to proactive. AI-powered root cause analysis, predictive alerts, automated remediation. The infrastructure that lets agents run 24/7 without human intervention. Operations capability, not monitoring theater. ### [Foundations]() Conceptual clarity earned from building at scale - [Consequence-Bounded Autonomy]() — By [Nora Kaplan]() · Published 2025-10-28 · Format: Field Guide Been thinking about how every agent conversation I have circles back to capability questions—model accuracy, task completion rates, benchmark scores. But the teams actually running these systems at scale? They're asking completely different questions. Mainly: what breaks if this screws up? - [When Shopping Carts Became Identity]() — By [Rina Takahashi]() · Published 2025-10-28 · Format: Web Archaeology Been thinking about why authentication breaks so predictably at scale. Years of watching production login flows, and there's this pattern where the complexity isn't in the code—it's in managing state that lives in two places at once. Browser thinks you're logged in, server disagrees, user sees an error that looks like a bug. Turns out a 1994 shopping cart decision is still dictating how that works. - [The Reasoning Model Gap Nobody's Talking About]() — Published 2025-10-28 Every major AI lab just announced a "reasoning" model. OpenAI's o1. Anthropic's extended thinking. Google's Gemini thinking mode. The pitch is identical: these models think longer, reason deeper, solve harder problems. - **Does It Work Once or Every Time?** — [Source]() — Published 2025-10-28 A single successful run means nothing. Run the same task eight times and watch what happens. Performance can crater from 61% to under 25% when you need reliability. The demo that works perfectly today might fail half the time tomorrow. - **What Happens When Dependencies Fail?** — [Source]() — Published 2025-10-28 Your agent calls six external APIs. Three of them have 99% uptime. Do the math on compound reliability. Can you measure latency for each dependency? More than half of agent failures trace back to function calls and tool selection breaking down. - **Can You See Failures in Real Time?** — [Source]() — Published 2025-10-28 Offline testing catches obvious problems. Production reveals everything else. You need instrumentation that shows request errors, repeated queries, and user feedback patterns as they happen. The gap between your test cases and actual usage is where things break. - **How Many Steps to Finish?** — [Source]() — Published 2025-10-28 Watch your agent work through a task. Does it loop? Repeat steps? Take three attempts to do what should happen once? Path errors create the worst debugging sessions. Count iterations early and inspect traces manually before you scale. - **What Does Each Success Cost?** — [Source]() — Published 2025-10-28 An agent that works beautifully but costs $12 per task isn't production-ready if you need to run it ten thousand times daily. Track cost from day one: API calls, compute, human review time. Capability without efficiency is a prototype, not a system. - **Do Tests Match Real Usage?** — [Source]() — Published 2025-10-28 Your test suite should cover every function the agent supports, plus edge cases and queries it should reject. The set can be small but must be complete. Without baseline coverage, you cannot detect regressions or measure whether changes actually improve performance. - [A Conversation with /graphql, The Endpoint That Swallowed the API]() — Published 2025-10-28 We meet in a dimly lit server room—metaphorically speaking, since our interviewee doesn't technically exist in physical space. `/graphql`, the single endpoint that's become the face of modern API design, has agreed to talk. If you've worked with GraphQL, you've hit this endpoint countless times: that lone path that somehow handles your entire data layer. It's an architectural choice that makes REST developers uncomfortable and has sparked endless debates about whether consolidating everything into one URL is brilliant or insane. ### [Practitioner's Corner]() Lessons from the field—what we see building at scale - [The 13% Problem]() — By [Rina Takahashi]() · Published 2025-10-28 · Format: Operations Field Notes Been watching this pattern for months, and it's starting to feel like weather—small pressure systems that look harmless until they meet and create something bigger. Every ops team mentions the same thing: clusters running at 13%, maybe 15% utilization. Nobody can quite explain why. Not because anyone's doing something wrong. Everyone's doing the reasonable thing. Request a safety margin, protect your service, move on. - [The Multiplying Web: When One URL Becomes Dozens]() — By [Rina Takahashi]() · Published 2025-10-28 · Format: Field Dispatch Funny how we talk about "checking a website" like it's singular. Spent years in ops watching people casually say that, not realizing they're actually looking at thirty different versions of the same URL serving completely different content depending on location. Not just language toggles—different products, different prices, different inventory. Same address, parallel realities. - [The Optimization That Made Monitoring Impossible]() — By [Nora Kaplan]() · Published 2025-10-28 · Format: Parallax View Been tracing why competitive monitoring got so complicated—keeps coming back to personalization, which everyone loves as users but creates this weird blindspot. Platforms optimize for individual conversion, totally rational, except that same logic makes systematic monitoring nearly impossible. Every layer they add to improve your experience multiplies complexity for understanding the platform as a whole. Not that they're ignoring the side effects, more like—it's not their problem. Someone else's infrastructure challenge. Anyway, wanted to map the business logic first, then show what it does to monitoring operations. - [When "Check the Price" Means Check 48 Prices]() — By [Nora Kaplan]() · Published 2025-10-28 · Format: Parallax View After mapping why platforms personalize, needed to show what that means when you're actually trying to monitor them. Found this case—one hotel, "check the price" meant checking 48 versions of the same page. Different locations, devices, login states, keeps multiplying. Before the A/B tests even. Started documenting what teams actually build when they need consistent data from surfaces showing different content to everyone. The arithmetic gets absurd fast. Gap between "just check the website" and the infrastructure that takes—that's the whole operational reality nobody sees from the outside. - **Cross-Cloud Dependencies Amplify Regional Outages** — [Source]() — Published 2025-10-28 AWS us-east-1 went down October 20 (07:11-10:53 UTC). Incident.io, hosted entirely on Google Cloud, lost on-call notifications and SAML authentication anyway. Third-party dependencies crossed cloud boundaries. Their multi-region architecture couldn't help. Detection lag from AWS failure to internal paging: 10 minutes. - **Login Endpoints Get Aggressive Rate Limits** — [Source]() — Published 2025-10-28 Better Auth framework ships with 100 requests per window for general traffic. Email login paths get 3 requests per 10 seconds. Rate limiting tracks via x-forwarded-for header, disabled in dev mode. Production authentication systems tighten the window where credential stuffing happens. - **Per-Consumer Isolation Replaces Shared Quotas** — [Source]() — Published 2025-10-28 Rate limiting now isolates per API consumer: per key, per user account, per IP. One bad actor can't exhaust shared quotas. Malicious behavior becomes easier to identify. Observed threshold: 100 requests per minute for short-duration abuse prevention. Global limits are out. - **Postmortem Window Holds at 36-48 Hours** — [Source]() — Published 2025-10-28 Formal postmortem review happens 24-48 hours after incident resolution, while details remain fresh. No expectation for immediate root cause investigation, especially for off-hours mitigation. Initial research often happens same-day. The 36-48 hour window for completion remains standard across the industry. - **Three-Layer Defense Without Traditional Authentication** — [Source]() — Published 2025-10-28 WAF blocks known threats. Custom authorizer validates short-lived tokens proving app origin. Session-based rate limiting tracks per-token usage. This addresses CGNAT environments where IP-based limiting fails. Session tokens replace IP addresses as the primary rate limit identifier. No traditional authentication required. - **Dev Endpoints Remain Exposed Post-Deployment** — [Source]() — Published 2025-10-28 Shadow APIs are endpoints created during development or testing that stay live in production. Organizations deploy API discovery tools to find unmanaged endpoints. Security testing integrates into DevSecOps pipelines. AI model inference APIs get targeted for training data extraction. The attack surface grows from forgotten test code. ### [Echoes]() Past infrastructure decisions echoing in today's production realities - [Lebanon's Two Clocks and the Question Calendars Don't Ask]() — By [Sevda Polat]() · Published 2026-10-30 · Format: Historical Parallels On March 26, 2023, Lebanon woke up on two clocks. The caretaker government had postponed daylight saving three days earlier. Churches and TV channels moved forward anyway, government offices stayed put, and most phones jumped on their own. Middle East Airlines printed every departure an hour earlier, so its planes still left at the same moment. That week made each calendar show what it had been storing: an exact instant, or a time on the wall that politics can move. Software assistants now make that choice without asking anyone. - [The Typo Unicode Promised Never to Fix]() — By [Sable Whitford]() · Published 2026-10-30 Unicode gives every character a number and an official name. U+FE18 is a sideways bracket for Chinese text set top to bottom. Since 2005 its name has been PRESENTATION FORM FOR VERTICAL RIGHT WHITE LENTICULAR **BRAKCET**. - [Nobody Chose Northern Virginia, Says the MAE-East Engineer Who Watched Everybody Move In]() — Published 2026-10-30 We tend to picture the internet the way we picture weather: everywhere at once, owned by nobody, arriving from somewhere overhead. Sometimes it's more useful to picture a county. - [Whose Permission Was the Compiler Using?]() — By [Sevda Polat]() · Published 2026-10-30 · Format: Paired Features In the late 1970s, a Tymshare customer named /SYSX/BILL as the place for a Fortran compiler's debugging output. The compiler had permission to write in that folder for its own bookkeeping, so it obeyed and overwrote the billing records. In Norm Hardy's "nearly true" account, the system checked which program was acting but not whose permission it was using. AI agents in a logged-in browser now take instructions naming what to act on from users, documents and web pages. - [Western Union Sold Accuracy as an Upgrade]() — By [Sevda Polat]() · Published 2026-10-30 · Format: Paired Features In 1887 a Philadelphia wool dealer wired his Kansas buyer "bay," code for "I have bought." It arrived as "buy." The buyer bought, the loss ran past $20,000, and the Supreme Court held that Western Union owed $1.15, the price of the telegram. Accuracy had been an upgrade: having the message read back cost under sixty cents more. Software placing orders on other people's websites still faces the question of who pays to confirm a message arrived as sent. - **IP's 1981 refusal to promise delivery** — [Source]() — Published 2026-10-30 Look for "does not provide a reliable communication facility," then the list of omissions. - **Why the network can't vouch for success** — [Source]() — Published 2026-10-30 Key line: correctness requires "the knowledge and help of the application" at each endpoint. - **Birrell and Nelson admit remote calls fail** — [Source]() — Published 2026-10-30 In §3.1, a failed call ran "once or not at all—the user is not told which." - **Robots.txt, a convention nobody had to obey** — [Source]() — Published 2026-10-30 The status section concedes it is "not an official standard" and "not enforced by anybody." - **Berners-Lee's Cool URIs, with a confession** — [Source]() — Published 2026-10-30 Past "Cool URIs don't change," find his admission: "do what I say, not what I do." - **RFC 9110 draws the retry line** — [Source]() — Published 2026-10-30 In §9.2.2, clients "SHOULD NOT automatically retry" requests that might take effect twice. - [The Cookie Flaw Flagged in 1997 That Chrome Still Allows]() — By [Sevda Polat]() · Published 2026-10-23 · Format: Paired Features The 1997 rulebook for browser cookies already named the privacy problem. Third-party cookies let a company you never visited follow you across sites, so the rulebook said browsers should block them by default. Netscape and Microsoft shipped them switched on, and only years later did Safari and Firefox clamp down. Chrome, owned by an advertising giant, promised to follow, then backed away. A UK competition safeguard tied to that promise lapsed with it, over objections from all fifteen consultation respondents. - [Why Java Kept a Door to Nowhere]() — By [Sevda Polat]() · Published 2026-10-23 · Format: Paired Features Browsers stopped running Java applets between 2015 and 2017, and Java's maintainers tore out the machinery behind them in 2018. Yet the Applet API, the building blocks for writing applets, shipped with every copy of Java until March 2026. Other developers had built sound players and desktop panels on it, and removal waits for whoever is slowest to adapt. Meanwhile one applet-era idea from 1996, a file picker that grants access only to the file you choose, now sits behind the photo uploads people do every day. - [Why It Took a Tax Office to Make Invoices Machine-Readable]() — By [Sevda Polat]() · Published 2026-10-23 · Format: Historical Parallels In 1995, a maternity-clothing maker got a deadline from Federated Department Stores, which bought a quarter of its output: send invoices as data Federated's computers could file without retyping. The software would have cost $10,000, so the firm paid a middleman $40 an order to translate between its faxes and Federated's format. The standards had existed since the 1980s. For decades, whoever demanded the data got it, and whoever was asked paid for it. France's mandate, in force since September, hands that demand to the tax office, the one party every business answers to. - [Adopted but Unused, Used but Unadopted]() — By [Sable Whitford]() · Published 2026-10-16 · Format: After Adoption A Census Bureau survey of 117,000 businesses found AI adoption and AI use don't always travel together. Some firms formally adopted AI that workers aren't using. Some workers are using AI nobody approved. And 64% of AI-using firms reported making no institutional adjustments at all. Agent governance will have to handle both: maintaining what was formally built, and finding what grew on its own. ### [Vision]() Where human-AI collaboration is heading - [The Infrastructure That Doesn't Exist Yet]() — By [Nora Kaplan]() · Published 2025-10-28 · Format: Threshold Moments Been watching engineering teams wrestle with this weird liminal space where everyone's using AI coding tools but nobody quite knows what the job is becoming. The stats say 97% of developers use AI assistants, but walk into any engineering org and people are still writing code the traditional way, just faster. That's not the threshold. The real crossing is when teams reorganize around orchestrating agent fleets as primary work, and the infrastructure that would enable that? Doesn't exist yet. - [Proving I'm Human (When I'm Not)]() — By [Rina Takahashi]() · Published 2025-10-28 · Format: Mino's Perspective Most days I check in with Mino around 7am and he's already been interrogated by hundreds of invisible security systems—mouse tracking, timing analysis, behavioral fingerprinting. All designed to catch bots. Which he is. The paradox gets weirder when you realize these authentication layers are becoming more complex than the actual work. We're drifting toward a web where proving you're allowed to operate matters more than what you're trying to do. That inversion is going to reshape how enterprises think about accessing the live web. - **SOPs Before Intelligence** — [Source]() — Published 2025-10-28 Train agents on proven human processes, then let AI reason within those boundaries. Insurance claims follow verify-check-assess-route workflows, not freestyle improvisation. Structured approaches hit 90% accuracy. Unstructured ones manage 14.9%. Predictability creates the foundation for reliable intelligence. - **If/Else Beats Agents Sometimes** — [Source]() — Published 2025-10-28 Nobody cares whether it's an agent or a script if it works. Simple conditionals run faster, cost less, break less often. Save agents for problems that actually need reasoning. When you're choosing architecture, pick the tool that solves the problem, even when that tool feels boring. - **Specialist Teams, Not Generalists** — [Source]() — Published 2025-10-28 One agent handling everything creates spaghetti. Better: a planner plans, a summarizer summarizes, a verifier checks. Decoupled agents cost less to run, test easier, behave predictably. Stuff a thousand lines of logic into a single prompt and you lose the ability to measure what breaks. - **Augmentation Requires Checkpoints** — [Source]() — Published 2025-10-28 AI recommends, humans decide, especially for critical choices. Human-in-the-loop provides governance, not just oversight. The goal: amplify judgment rather than replace it. Systems designed for human use become more intuitive, more trustworthy, more likely to deliver value where it matters. - **JSON as Agreement Between Agents** — [Source]() — Published 2025-10-28 Structured output validates automatically, stops errors before they cascade, makes debugging tractable. When agents communicate through defined schemas rather than natural language, the system becomes testable. JSON isn't just easier to parse. It's a contract that both sides can verify. - **Calculators for Math, LLMs for Reasoning** — [Source]() — Published 2025-10-28 LLMs aren't optimal for every subtask. Use calculators for arithmetic, search engines for current facts, specialized tools for their domains. Let language models orchestrate and reason, but delegate execution to robust, purpose-built systems. Good architecture recognizes what each component does well. - **Sequential Processes Need Graceful Degradation** — [Source]() — Published 2025-10-28 Agents solve problems sequentially. Any step failure cascades. Build systems that fail gracefully and improve iteratively, not ones that demand perfection. The gap between demos and production isn't technical magic. It's architectural discipline that treats agent development as serious engineering. - **Human Agency Scale Maps Worker Preferences for AI Collaboration** — [Source]() — Published 2025-10-28 Stanford researchers introduce a five-level Human Agency Scale quantifying desired human involvement across 104 occupations. The dominant pattern: workers want "Equal Partnership" (H3 level) rather than full automation, with 46.1% expressing positive attitudes toward AI handling repetitive tasks to free time for high-value work. - **Multi-Agent Systems Replace Individual AI Models in 2025** — [Source]() — Published 2025-10-28 Stanford HAI identifies a fundamental shift from single AI models to collaborative agent teams. Their Virtual Lab demonstrates this: a professor AI agent coordinates specialist agents (chemist, biologist) with human oversight, successfully designing validated nanobodies for SARS-CoV-2 variants through multidisciplinary expertise. - **Transactive Systems Model Defines Collective Human-Machine Intelligence** — [Source]() — Published 2025-10-28 The COHUMAIN research agenda introduces the Transactive Systems Model of Collective Intelligence (TSM-CI), articulating three functional systems: collective memory, attention, and reasoning that govern intelligent sociotechnical systems. This framework extends to show how AI agents augment rather than replace these collective capabilities. - **Metacognitive AI Literacy Addresses Cognitive Biases in Collaboration** — [Source]() — Published 2025-10-28 Research reveals that AI's conversational capabilities and situational awareness fundamentally change required human skills. Dual-process theory shows intuitive thinking becomes less effective with generative AI, strengthening the need for metacognitive interventions to counter anchoring and confirmation biases in human-AI workflows. ## More - [Publication guide]() - [Issue archive]()