--- title: "Issue 16" canonical: "https://current.tinyfish.ai/issue/16" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/16" issue: 16 description: "Infrastructure knowledge lives in the strangest places—TLS handshakes, normalized-away signals, morning rituals that dissolve. This week: the production gaps between what systems promise and what actually works." authors: ["Sable Whitford","Nora Kaplan","Rina Takahashi"] published: 2026-02-16 updated: 2026-11-05 articles: 94 --- # 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 16 Infrastructure knowledge lives in the strangest places—TLS handshakes, normalized-away signals, morning rituals that dissolve. This week: the production gaps between what systems promise and what 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. - [The Runnable Paper Needs an Identity Report]() — Published 2026-10-30 Paper2Agent, published in Nature in September, turns research papers into AI agents you can run. Hand over your data and the method executes, with no rebuilding it from prose. The authors want papers to carry an "agent availability" section saying whether a runnable version exists. - [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. - [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. - [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 - [Why Agent Platforms Still Deploy Humans]() — By [Nora Kaplan]() · Published 2026-02-11 · Format: The Signal Salesforce is hiring Forward Deployed Engineers at a pace that saw job postings surge 800% in nine months. OpenAI's Frontier platform launched with FDEs who embed with enterprise teams through deployment. These are technical experts who stay through production, working alongside teams to get agents operational. - [Five Agent Trajectories Worth Watching Through Mid-2026]() — Published 2026-02-11 Several patterns from recent weeks deserve attention as they develop over the coming months. - **Graph-Based Agent Memory: From Logs to Topology** — [Source]() — Published 2026-02-11 Survey analyzing how graph structures transform agent memory from passive logs into topological models that preserve relationships. Covers extraction, storage, retrieval, and evolution techniques with open-source libraries and benchmarks builders can actually use. - **Memory in the Age of AI Agents: Unified Taxonomy** — [Source]() — Published 2026-02-11 Forty-seven authors consolidate fragmented agent memory research into unified taxonomy covering forms, functions, and dynamics. Distinguishes agent memory from LLM memorization across when learning happens, where it lives, and how it's governed. - **Agent Workflow Memory: Learning Reusable Task Patterns** — [Source]() — Published 2026-02-11 Enables agents to induce commonly reused workflows from past experiences and apply them to new situations. Tested on Mind2Web and WebArena benchmarks covering 1000+ tasks from 200+ domains, with significant improvements in success rates. - **A-MEM: Self-Organizing Memory Using Zettelkasten Principles** — [Source]() — Published 2026-02-11 Memories actively generate their own contextual descriptions, form meaningful connections, and evolve structure over time. Inspired by proven human knowledge management system, exhibits agency at fundamental level through autonomous evolution of memory structure. - **Fault Tolerance as Built-In Primitive** — [Source]() — Published 2026-02-11 Temporal and Restate externalize agent state into systems designed for recovery. Processes crash mid-workflow and resume exactly where they stopped. No reimplemented retries, no lost progress. Reliability becomes infrastructure, not hand-coded logic that breaks under load. - **Kernel-Level Sandboxing for Agent Code** — [Source]() — Published 2026-02-11 Google's Agent Sandbox uses gVisor to isolate LLM-generated code at the kernel level. Sub-second latency for fully isolated workloads. Kubernetes gets a native primitive for running untrusted agent code safely in production environments. - **Zero-Copy Forks for Isolated Testing** — [Source]() — Published 2026-02-11 Agentic Postgres provisions database forks in seconds using copy-on-write storage. Agents spin up isolated environments with full production data without duplicating costs. Testing becomes lightweight infrastructure instead of expensive replication. - **OpenTelemetry Conventions for Agent Traces** — [Source]() — Published 2026-02-11 Standardized semantic conventions define how to trace tasks, actions, agents, and memory across frameworks. Microsoft and Cisco are collaborating on consistent telemetry. Works whether you're using OpenAI, Anthropic, or LangChain. Cross-platform observability without vendor lock-in. - **Semantic Context for Agent Diagnostics** — [Source]() — Published 2026-02-11 Knowledge graphs connect services, teams, datasets, and alarms with meaningful relationships. When alerts fire, agents navigate the graph to pull relevant metrics, logs, and traces. Understanding context instead of brute-forcing through raw telemetry. - **Real-Time Tracking for Agent Economics** — [Source]() — Published 2026-02-11 Inference costs plummet but usage explodes. Enterprises need infrastructure to track agent spending in real time. Paid and similar platforms address the gap between per-seat pricing and continuous inference economics. - **LangChain Survey: Everyone Watches Agents, Half Never Test Them** — [Source]() — Published 2026-02-11 1,300 practitioners admit the gap: 89% add observability, 52% skip systematic evaluation. Monitoring without validation is just expensive logging. - **Braintrust: Why Agent Evaluation Needs Infrastructure, Not Scoring** — [Source]() — Published 2026-02-11 Multi-step agents break prompt-response testing. Practical framework for why tracing becomes foundational, not optional, for production evaluation. - **Microsoft: 29% of Fortune 500 Agent Usage Is Shadow AI** — [Source]() — Published 2026-02-11 Deployment velocity outpaces governance infrastructure. The real number: organizations building agent capabilities faster than they can establish controls. - **Gravitee: 88% Hit Agent Security Incidents in Production** — [Source]() — Published 2026-02-11 Security failures moved from theoretical risk to baseline reality. Survey of 900+ executives shows incidents are infrastructure problems now. - **Cloud Security Alliance: 78% Still Use API Keys for Agents** — [Source]() — Published 2026-02-11 Only 22% treat agents as independent identities. Shared credentials and API keys reveal how far behind identity architecture lags deployment. - **Anthropic Engineering: How We Actually Evaluate Our Agents** — [Source]() — Published 2026-02-11 Rare practitioner view from builders shipping production systems. Dedicated eval teams plus domain experts owning task definitions. The model that works. ### [Foundations]() Conceptual clarity earned from building at scale - [When Agents Can't Find What's Right in Front of Them]() — By [Rina Takahashi]() · Published 2026-02-11 · Format: Field Guide A web scraper pulls product prices from a retail site every morning. One day it returns empty results. The prices are still there, displayed correctly on the page—customers see them, competitors see them. But the agent can't find them. Teams spend days checking pipelines and validating formats. Something broke, something most debugging approaches miss entirely. - [Headless Browsers at Scale: Where Speed Determines Economics]() — By [Rina Takahashi]() · Published 2026-02-11 · Format: Tools in Context The infrastructure lead switches 10,000 daily hotel price checks from headful to headless browsers. Monthly bill: $4,500 to $1,500. Tasks complete 2-15x faster. Same server handles three times the concurrent sessions. For teams running automation at volume, headless execution determines how many servers you need and what they cost. Then anti-bot systems evolve to catch the missing rendering pipeline, the absent plugins, the suspicious timing patterns. Engineering around detection becomes its own cost, one that sometimes exceeds the infrastructure savings. - [When Headful Browsers Justify Their Resource Overhead]() — By [Rina Takahashi]() · Published 2026-02-11 · Format: Tools in Context The pricing monitoring system ran cleanly for three months—15,000 hotel properties checked daily, data flowing into automated pricing decisions. Then Booking.com updated their anti-bot system. Success rate dropped from 94% to 31% within 48 hours. Missing pricing data meant stale dashboards. Stale dashboards meant the system underpriced 400 properties based on weekend rates that were four days old. The revenue team calculated the loss. The cost of being wrong exceeded the cost of running more expensive infrastructure by a factor of eight. - [Platform Vendors Launch Before Production Proves Out]() — Published 2026-02-11 Three major vendors launched agent management platforms in February 2026. OpenAI released Frontier on February 5. Salesforce expanded Agentforce throughout January and February. AWS grew Bedrock AgentCore. Gartner called these platforms "the most valuable real estate in AI." - **Does It Handle Corrupted Files?** — [Source]() — Published 2026-02-11 Demos use clean inputs. Production delivers corrupted PDFs, malformed JSON, and queries that make no sense. The tool that chokes on the first weird file isn't a tool. It's a liability waiting to page you at 2 AM. - **Can It Run 10,000 Tests Simultaneously?** — [Source]() — Published 2026-02-11 Sequential testing is a luxury production can't afford. You need parallel execution across distributed resources, not a queue that takes hours. If the tool can't shard tests and run them concurrently, you're building a bottleneck into your pipeline. - **Does QA Actually Match Production?** — [Source]() — Published 2026-02-11 Code that works in staging and fails in production is expensive. Not just frustrating. Your QA environment needs the same hardware, network topology, and configuration drift that production has. Otherwise you're testing a fantasy version of your system. - **Are Integrations Native or Duct-Taped?** — [Source]() — Published 2026-02-11 Bolted-on integrations break. Native ones fade into infrastructure. Check whether the tool connects to your actual systems—GitHub, Datadog, PagerDuty—without custom glue code. If you're writing adapters, you're maintaining two systems instead of one. - **How Fast Can You Undo This?** — [Source]() — Published 2026-02-11 Rollback sounds simple until you try it with database migrations and dependent services. The tool should make reverting trivial, not a multi-step procedure requiring three teams. If rollback takes longer than the deploy, something's wrong with your architecture. - **Will This Company Survive Your Contract?** — [Source]() — Published 2026-02-11 Tool selection is a three-year bet minimum. Check how long they've existed, whether they're actively shipping updates, and if their business model makes sense. Migrating off a dead vendor is worse than choosing wrong in the first place. - [An Interview with the Acceleration Curve That Judges Your Humanity]() — Published 2026-02-11 We're sitting in what I can only describe as a visualization lab. Walls covered in projected trajectories, each one a ghost of cursor movement past. Some flow like water. Others snap corner to corner like they're being drawn with a ruler. ### [Practitioner's Corner]() Lessons from the field—what we see building at scale - [What Got Normalized Away]() — By [Nora Kaplan]() · Published 2026-02-11 · Format: Field Dispatch The competitor tracking dashboard showed stable availability for three months, then a 12% price jump overnight. The data quality engineer pulled last quarter's manual collection logs to compare. Same product, same retailer, two weeks before the spike: "Limited stock—2-3 units available." This quarter's agent-collected entry for the same listing: "In stock." Three months of records told the same story. Something in the automation was smoothing away signals that preceded price changes. The data had gotten cleaner. - [What You Lose When You Remove the Screenshots]() — By [Rina Takahashi]() · Published 2026-02-12 · Format: Builder Profiles A login button moves from top-right to bottom-left during a site redesign. The HTML is identical—same element, same attributes—but your agent can't find it anymore. The spatial cues that signaled "this matters" or "wait here" vanish with the visual information when you strip the page down to text. - [TLS Fingerprinting Creates 47x Success Rate Gap]() — Published 2026-02-11 Standard Python requests library achieves roughly 2% success rate against Amazon's 2026 anti-bot system. Switch to curl_cffi, which mimics browser TLS fingerprints, and success climbs to 94%. That's a 47-fold improvement from addressing a single technical layer most developers never see. - **False Positives Terminated 3% of Railway** — [Source]() — Published 2026-02-11 Railway's automated abuse detection incorrectly terminated legitimate workloads February 11, sending unexpected SIGTERM signals to active databases. New anti-fraud ruleset delivered false positives. Postgres and MySQL instances went offline across roughly 3% of platform services. - **USPS Caps New APIs at 60 Hourly** — [Source]() — Published 2026-02-11 Legacy Web Tools retired January 25, 2026. Replacement APIs enforce default rate limits of 60 requests per hour. Higher thresholds exist but USPS doesn't publish available tiers or approval criteria. High-volume shipping operations now operate blind. - **Atlassian Switches to Points-Based API Quotas** — [Source]() — Published 2026-02-11 Starting March 2, 2026, Jira and Confluence REST and GraphQL requests consume points based on work performed: data returned, operations triggered. Hourly quotas allocated per-tenant, per-app. No more simple request-per-second caps. - **Microsoft Patched 6 Actively Exploited Zero-Days** — [Source]() — Published 2026-02-11 February 2026 Patch Tuesday addressed 58 vulnerabilities including 6 actively exploited in the wild. CVE-2026-21533 allows attackers to gain SYSTEM privileges through Remote Desktop Services. CVE-2026-21525 denial-of-service exploit found in public malware repository. - **Azure West US Took 20 Hours Recovery** — [Source]() — Published 2026-02-11 Power interruption starting February 7 at 07:58 UTC caused broad infrastructure reachability loss. As power stabilized, storage and compute components failed to return to healthy state. Cascading failures across dependent components. Full resolution February 8 at 04:24 UTC. - **Integration Issues Kill Agents, Not LLMs** — [Source]() — Published 2026-02-11 Three leading causes: bad memory management (dumb RAG), broken I/O (brittle connectors), and polling-based architectures. Polling wastes 95% of API calls, burns through quotas, never achieves real-time responsiveness. Webhooks or event-driven systems only. - **Railway's Fraud Prevention Incident: When Security Breaks Production** — [Source]() — Published 2026-02-11 Fraud detection sent SIGTERM to 3% of workloads. Postmortem traces how security tooling cascaded into production failures. - **The 8 Fallacies Meet AI: Distributed Systems in 2026** — [Source]() — Published 2026-02-11 Those classic fallacies now apply to LLM inference and agentic chains. Includes actual latency numbers from production systems. - **AI Agent Observability: The Verifiability Problem Nobody's Solving** — [Source]() — Published 2026-02-11 Organizations deploy digital labor they cannot inspect or govern. ClickHouse architect explains why observability lags deployment velocity. - **Web Scraping Infrastructure: Where Teams Actually Spend Money** — [Source]() — Published 2026-02-11 Nearly half build internal scrapers, most saw infrastructure costs rise. Survey data reveals what scraping at scale demands. - **Cursor AI Architecture: How AI-Powered IDEs Handle Multi-File Reasoning** — [Source]() — Published 2026-02-11 Task scheduling, memory optimization, context processing. Technical breakdown of how production AI development tooling actually works. - **Docker BuildKit 0.17: Multi-Architecture Builds and Content-Addressable Storage** — [Source]() — Published 2026-02-11 Cross-platform capabilities, incremental builds, cache optimization. BuildKit's architecture matters for modern CI/CD at scale. ### [Echoes]() Past infrastructure decisions echoing in today's production realities - [The Authentication System That Doesn't Call Home]() — By [Rina Takahashi]() · Published 2026-02-11 · Format: Infrastructure Inflections India processes 2.5 billion authentication transactions monthly through a system that verifies identity locally. When a bank confirms who you are using Aadhaar, verification happens on their own systems—a digitally signed XML file, validated against a public key, entirely within the service provider's infrastructure. - [When HTTP/2 Server Push Vanished Without Replacement]() — Published 2026-02-11 HTTP/2 Server Push arrived in 2015 with a straightforward promise: servers could send resources before browsers asked for them. The specification included the feature. Browsers implemented it. Then servers hit a wall. They couldn't determine what clients already had cached, which meant pushing resources that browsers didn't need. Bandwidth waste followed. Performance regressions appeared. - **Cloud Bills Force Strategic Repatriation** — [Source]() — Published 2026-02-11 Eighty-six percent of organizations are pulling workloads back from public cloud. The economics that worked at prototype scale break at production volume. Hybrid infrastructure isn't compromise—it's companies learning what actually costs less to run. - **SaaS Proliferation Hits Economic Limit** — [Source]() — Published 2026-02-11 Companies average 106 SaaS applications with $21M in annual waste. Over 2,600 M&A deals in 2025 signal the turn. Point solutions that multiplied during growth now consolidate under platform pressure. The pattern always ends here. - **Automation Creates Different Work, Not Less** — [Source]() — Published 2026-02-11 AI handles routine tasks, then workers become system stewards managing edge cases. Manufacturing firms adopting AI see productivity decline. The work doesn't disappear. It transforms into something harder—oversight of systems that fail in unexpected ways. - **Kubernetes Spawns Services to Hide Itself** — [Source]() — Published 2026-02-11 Developers spend seventy percent of time on infrastructure. Eighty-eight percent of teams report higher costs. AWS EKS and managed services emerge to abstract the complexity Kubernetes created. Every infrastructure layer eventually generates its own abstraction market. - **AI Inference Shatters Power Planning** — [Source]() — Published 2026-02-11 After twenty years of flat electricity demand, data centers will consume twelve percent of US power by 2030. AI workloads break assumptions built over decades. Infrastructure designed for predictable growth hits a wall when workload patterns fundamentally change. - **Binary Decisions Give Way to Workload Selection** — [Source]() — Published 2026-02-11 Only eight percent of organizations fully exit cloud. The rest manage hybrid portfolios, moving some workloads while expanding others. Infrastructure choices become optimization problems, not religious commitments. The all-or-nothing phase always gives way to selective strategy. - **REST Dissertation Defined Modern API Architecture** — [Source]() — Published 2026-02-11 Roy Fielding's 2000 doctoral dissertation introduced REST for distributed hypermedia systems. Built for web-scale document linking, not APIs, it became the foundation for modern web service communication. Most implementations claiming REST principles actually diverge from Fielding's original architectural constraints. - **SOAP Standardization Created Enterprise Web Services Complexity** — [Source]() — Published 2026-02-11 W3C standardized SOAP between 2000-2003 as XML-based protocol for web services. Rigid standards, verbose messages, and WS-* extensibility dominated early 2000s API design. Developer frustration with SOAP's complexity triggered mass migration to REST, fundamentally reshaping web architecture patterns. - **HTTP/2 Solved Application-Level Performance While Creating New Problems** — [Source]() — Published 2026-02-11 Published as RFC 7540 in May 2015, HTTP/2 introduced binary framing, multiplexing, and header compression to address HTTP/1.1's performance limitations. Eliminated application-level head-of-line blocking but created packet-level blocking at TCP layer, now driving HTTP/3 development. - **Certificate Transparency Made TLS Certificate Issuance Public** — [Source]() — Published 2026-02-11 Published as RFC 6962 in June 2013 after DigiNotar's 2011 compromise, Certificate Transparency created append-only public logs for all TLS certificates. Google engineers designed cryptographic proof systems to detect mis-issued certificates, fundamentally changing how certificate authority trust operates. - **Netflix's 2009 Microservices Pattern Copied Without Context** — [Source]() — Published 2026-02-11 Teams adopted the architecture without the scale problems it solved. Now 42% quietly rebuild monoliths from the pieces. - **Roy Fielding's REST Principles Meet Facebook's GraphQL Response** — [Source]() — Published 2026-02-11 Two architectural philosophies from different web eras collide. Teams now maintain hybrid systems neither designer imagined would coexist. - **Chrome's Spectre Defense Architecture Faces AI Browser Agents** — [Source]() — Published 2026-02-11 Site isolation protected against CPU vulnerabilities in 2018. Autonomous agents accessing sensitive data reveal gaps the architecture never anticipated. - **COBOL Optimization Choices From Expensive Computing Era** — [Source]() — Published 2026-02-11 Mainframe decisions made when cycles cost real money. AI analysis tools finally make understanding those architectural choices economically feasible. - **Authentication Logic Moved Client-Side For User Experience** — [Source]() — Published 2026-02-11 Business logic followed authentication to browsers for speed. The attack surface that created now requires rebuilding security from foundation. - **Global Bank's $15 Million Annual Maintenance Budget** — [Source]() — Published 2026-02-11 Optimization decisions from different computing economics compound over decades. Infrastructure spending now consumes capacity that could drive innovation forward. ### [Vision]() Where human-AI collaboration is heading - [The Morning Check That Stopped Happening]() — By [Nora Kaplan]() · Published 2026-02-11 · Format: Threshold Moments The verification script sits three commands up in the terminal history. Tuesday's run, or maybe last week's. The analyst scrolls past it to start Wednesday's work—competitive pricing analysis, market intelligence updates. The coffee is still hot when the absence registers. When did the morning check become optional? - [When Continuous Workloads Break Cloud Economics]() — By [Rina Takahashi]() · Published 2026-02-11 · Format: Infrastructure Economics Cloud economics assumed workloads would spike and scale to zero. Training runs, batch jobs, traffic surges—all episodic. By early 2026, inference consumed over 55% of AI infrastructure spending. Inference runs continuously, serving requests 24/7. When investors demand ROI in six months and infrastructure costs must grow slower than earnings, the continuous nature of inference workloads exposes limits in the elasticity model. The economics that worked for episodic compute face different constraints. - [Why Cost Predictability Became a Reliability Problem]() — By [Rina Takahashi]() · Published 2026-02-11 · Format: Infrastructure Economics A system running at 99.9% uptime with costs swinging 40% quarter-to-quarter creates a reliability problem when CFOs need to model infrastructure spend. The dashboard shows green. The P&L shows chaos. Continuous inference workloads—the ones dominating AI infrastructure spending—revealed a gap: cost predictability matters as much as operational uptime. Reliability architecture had to account for structural dependencies that determine whether you can afford to keep running at the scale customers need. - [An Interview with Mino About the Web's Invisible Rooms]() — Published 2026-02-11 We're meeting Mino for the ninth time, and they're practically vibrating with excitement about something new. "Okay, so I need to tell you about this thing I keep running into," they start before I've even asked a question. "There are parts of websites that technically don't exist. Except they do. Except they don't. It's wild." - **Does Faster Always Mean Better** — [Source]() — Published 2026-02-11 Two days, fifteen iterations. Five days, two iterations. Same result. Traditional efficiency metrics break here. Maybe what we're measuring isn't what matters anymore. Or maybe we're watching agents optimize for the wrong thing while we count the wrong numbers. - **Where Coordination Costs Exceed Intelligence Gains** — [Source]() — Published 2026-02-11 Multi-agent inquiries jumped 1,445% in eighteen months. Coordination overhead grows exponentially. There's a breaking point where adding another agent subtracts more value than it creates. Nobody knows where that line is until they've crossed it. - **Why Some Organizations See 192% Returns** — [Source]() — Published 2026-02-11 While others abandon projects entirely. Same technology, radically different outcomes. Change management and governance maturity explain the gap better than capability differences. Are we watching temporary growing pains or permanent competitive restructuring? The divergence keeps widening. - **Rebuild Foundations or Adapt Agents** — [Source]() — Published 2026-02-11 Forty percent of agentic projects may fail by 2027. Not because agents aren't capable. Because infrastructure lacks real-time execution, modern APIs, modular architecture. Should organizations rebuild first or force agents to work within constraints? The choice determines velocity. - **Quality Versus Speed Versus Autonomy** — [Source]() — Published 2026-02-11 Pick two. Quality remains the top barrier. Latency emerges second. Multi-step agents deliver better outputs but slower responses. Customer-facing work demands speed. Complex decisions require thoroughness. Context determines which tradeoff matters. The tension doesn't resolve. - **When Experiments Become Mission Critical** — [Source]() — Published 2026-02-11 Two-thirds of organizations experiment with agents. Fewer than one in four reach production. The gap between pilot and platform isn't technical. It's definitional. When does an experimental system require enterprise-grade reliability? Usually right after it becomes indispensable. - **LLM Agent Maturity Model for Research Collaboration** — [Source]() — Published 2026-02-11 Nature Computational Science maps how research agents mature from simple task automation into genuine scientific collaborators. The framework tackles a thorny question: as AI handles more of the analytical workflow, what exactly should human scientists contribute? SciSciGPT demonstrates the progression while surfacing tensions around reproducibility and intellectual ownership. - **Capability Asymmetry Drives Human-AI Complementarity** — [Source]() — Published 2026-02-11 Taylor & Francis research quantifies what makes human-AI teams actually work. When AI deliberately fails where humans excel and succeeds where they struggle, complementarity jumps from 24% to 93% of team performance. The finding gives us a rigorous framework: asymmetric capabilities, not just different information, create collaborative value. - **AI Explanations Paradoxically Increase Over-Reliance** — [Source]() — Published 2026-02-11 University of Washington researchers found something unsettling: AI-generated explanations made people more likely to accept algorithmic recommendations, particularly when they should have rejected them. Transparency backfires. The better AI explains itself, the less critically humans evaluate its suggestions. Convenience trumps skepticism. - **Meta Agents Automatically Design Superior Agent Systems** — [Source]() — Published 2026-02-11 ICLR 2025 research shows meta agents writing code for progressively better agents across coding, science, and math tasks. The invented systems outperform hand-designed ones and transfer across domains. Programming languages being Turing complete means any possible agentic system could theoretically emerge through this automated discovery process. ## More - [Publication guide]() - [Issue archive]()