--- title: "Foundations" canonical: "https://current.tinyfish.ai/issue/12/foundations" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/12/foundations" issue: 12 section: "foundations" description: "Conceptual clarity earned from building at scale" authors: ["Rina Takahashi","Nora Kaplan"] published: 2026-01-18 updated: 2026-01-15 articles: 11 --- # 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 12 — Foundations Conceptual clarity earned from building at scale ### What Building Teaches Builders - [The Misidentification That Breaks Production Contracts]() — By [Rina Takahashi]() · Published 2026-01-14 · Format: Field Guide Your web agent extracts hotel prices perfectly for three months, then starts returning subtly different data structures—not errors, just variations. Your competitor's system breaks entirely when a site redesigns, but fails predictably. Same automation label, opposite failure modes. The industry treats deterministic automation and probabilistic agents as interchangeable solutions. When you're operating across thousands of sites with real production dependencies, that misidentification becomes operationally expensive. It determines what reliability guarantees are architecturally possible. ### Tools & Techniques - [Why 30,000 Teams Choose the Slower Parser]() — By [Nora Kaplan]() · Published 2026-01-14 · Format: Tools in Context The most popular HTML parser in web automation is five times slower than its alternatives. Cheerio dominates with 30,000 stars and 11 million weekly downloads while faster options sit in obscurity. Teams aren't missing the benchmarks. They're optimizing for the real bottleneck: hours spent maintaining extraction logic across thousands of sites that change without warning, not milliseconds spent parsing HTML strings. - [When Five Milliseconds Costs Three Hours]() — By [Nora Kaplan]() · Published 2026-01-14 · Format: Tools in Context At ten thousand pages, parsing speed is a benchmark number. At a million pages, it's three hours of compute time. The gap between fast and slow HTML parsers stops being academic when you're processing massive volumes daily. Teams reaching for node-html-parser aren't chasing optimization for its own sake. They've hit the scale where that 10-millisecond difference per page determines whether their pipeline meets its throughput requirements. ### Pattern Recognition - [Agents Deploy Faster Than Security Can Follow]() — Published 2026-01-14 Enterprises are shipping agents into production without the security infrastructure to contain them. Exabeam just launched Agent Behavior Analytics because traditional monitoring doesn't work for this new digital workforce. Chinese cyberspies already weaponized Claude Code in September attacks. Only 8.6% of companies actually have agents in production, yet 72% of S&P 500 firms now flag AI as material risk. ### Questions Worth Asking The questions that matter most aren't the ones vendors answer in demos. They're the ones that surface three months into production, when the initial excitement has worn off and you're dealing with the actual system. We've watched enough teams move from prototype to production to recognize the patterns. The questions that predict success aren't about capabilities. They're about operational reality. What happens when your agent encounters the messiness of real users, real data, and real infrastructure. These are the questions we ask ourselves when evaluating approaches. - **What's Your Tool Calling Failure Rate?** — [Source]() — Published 2026-01-14 Tag: Tool Reliability Production systems see tool calling fail 3-15% of the time. Does your architecture handle that gracefully, or does every failure cascade into user-visible errors? You need to build for failure from day one, not discover it three months in. - **Are You Drowning Your Agent in Context?** — [Source]() — Published 2026-01-14 Tag: Context Engineering Most production failures are context failures, not model failures. Models with 200K token windows still degrade when critical information gets buried in noise. Just-in-time retrieval beats pre-loading everything. Stop treating context windows like infinite resources. - **What Changes Between Demo and Production?** — [Source]() — Published 2026-01-14 Tag: Production Gap Demos lack cost-per-query budgets, p99 latency targets, and security postures guarding against prompt injection. A demo for five users is nearly free. An agent for 500 employees making thousands of queries daily can shock you with API bills. - **How Many Steps Before Human Intervention?** — [Source]() — Published 2026-01-14 Tag: Human Intervention 68% of production agents execute at most 10 steps before requiring human oversight. If your architecture assumes fully autonomous operation, you're designing for the exception. Build for supervised autonomy from day one. - **Who Maintains This After Launch?** — [Source]() — Published 2026-01-14 Tag: Day-2 Operations When backend APIs undergo breaking changes, does your architecture insulate the agent or require prompt rewrites? The engineering discipline required for maintenance often exceeds the initial deployment effort. Agents degrade constantly without active management. - **Can Your Data Architecture Support Agents?** — [Source]() — Published 2026-01-14 Tag: Integration Reality Traditional enterprise systems weren't designed for agentic interactions. Most agents still rely on APIs and conventional data pipelines, creating bottlenecks. Nearly half of organizations cite data searchability as a challenge. Your data architecture determines what's possible. ### In Dialogue With Complexity - [An Interview with Port 9222]() — Published 2026-01-14 We're sitting in what 9222 insists is "not really a place, more like a listening state," though it feels remarkably like a small, dimly lit server room. Chrome 136's upcoming security changes have made this conversation timely, if slightly awkward. Like interviewing someone who's just been told they need to move to a gated community. ## Other sections in Issue 12 - [Daily Brief]() - [Market Pulse]() - [Practitioner's Corner]() - [Echoes]() - [Vision]() ## More - [Publication guide]() - [Issue archive]()