--- title: "Foundations" canonical: "https://current.tinyfish.ai/issue/14/foundations" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/14/foundations" issue: 14 section: "foundations" description: "Conceptual clarity earned from building at scale" authors: ["Rina Takahashi"] published: 2026-02-02 updated: 2026-01-29 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 14 — Foundations Conceptual clarity earned from building at scale ### What Building Teaches Builders - [Three Territories of Web Automation Reliability]() — By [Rina Takahashi]() · Published 2026-01-28 · Format: Field Guide Your automation runs flawlessly through a hundred test cycles—clean data, zero errors, perfect repeatability. Scale to production and three weeks later you're explaining why revenue dashboards went blank overnight. The script still works. Something else broke. At TinyFish, we see this constantly: teams prototype in one operational reality, then deploy into another. By the time they realize the mismatch, downstream systems already depend on infrastructure that can't deliver what production actually demands. ### Tools & Techniques - [Selective Preservation Works When You Know What Matters Before You Look]() — By [Rina Takahashi]() · Published 2026-01-28 · Format: Tools in Context Extract exactly what you need: the price, the availability flag, the product title. Nothing more. This approach assumes you can predict what matters before you look—that sites behave consistently enough to target specific data points without preserving surrounding context. For some sites, that assumption holds. Production reveals which ones, and the answer surfaces only after you've run thousands of extractions across months of operation. - [Comprehensive Preservation for Context You Can't Predict]() — By [Rina Takahashi]() · Published 2026-01-28 · Format: Tools in Context Preserve everything first. Settled HTML after JavaScript execution, screenshots showing visual layout, network traces capturing the full request chain. Extract your data from that preserved state, then keep the state itself. Storage costs multiply, but teams choose this because the web surprises you—authentication flows that shift by region, bot detection that serves different content, layout changes that break extraction logic you can't debug without evidence. ### Pattern Recognition - [Companies Redesign Work Around AI Agents as Teammates]() — Published 2026-01-28 Organizations are restructuring hiring and roles to accommodate AI agents as operational colleagues. Entry-level hiring alterations jumped from 18% to 64% in one quarter. HCLTech trained 116,000 employees while creating orchestration engineers who shape agent execution and responsible AI engineers who build guardrails. ### Questions Worth Asking Most teams ask the wrong questions before deploying agents. They focus on capability demonstrations instead of operational reality. The gap shows up fast. We've watched this pattern repeat: frameworks get stripped out, integration takes three times longer than building, maintenance compounds until someone gets paged at 2am. The questions that actually predict success come from teams who've operated agents at scale and learned what breaks. These six questions cut through vendor promises to what matters in production. They're born from observable failures: the infinite loops, the drift nobody noticed, the SLAs you can't actually meet. Ask them before you deploy, not after. - **Will This Become Invisible Infrastructure?** — [Source]() — Published 2026-01-28 Tag: Framework Dependency Production teams eventually strip out frameworks. The pattern repeats: what ships demos creates maintenance burden at scale. You need infrastructure that fades into the background, not another system demanding constant attention and specialized knowledge to operate. - **How Wide Is Your Demo Gap?** — [Source]() — Published 2026-01-28 Tag: Deployment Reality Controlled environments hide what breaks under load. Agents enter infinite loops, produce inconsistent outputs, fail catastrophically when traffic patterns shift. The distance between "works on my machine" and "handles Tuesday afternoon reliably" predicts your first month in production. - **Can You Write Contracts on This?** — [Source]() — Published 2026-01-28 Tag: SLA Feasibility Traditional SLAs assume deterministic behavior. Your agent chooses different approaches for identical problems. You're either measuring adherence after failures happen or maintaining guarantees in real time. Probabilistic systems struggle with contractual commitments for a reason. - **Does Integration Dwarf Development Time?** — [Source]() — Published 2026-01-28 Tag: Integration Overhead Hooking agents into existing systems consistently takes longer than building them. Legacy infrastructure lacks modern APIs. Production agents live inside complex ecosystems with databases, tools, and workflows that weren't designed for probabilistic behavior. Budget accordingly. - **Can You Debug Without Full Traces?** — [Source]() — Published 2026-01-28 Tag: Observability Requirements You can't fix what you can't see. Production issues require context: what the agent was thinking, which tools it called, what errors occurred. Systems that don't log full interaction traces, reasoning chains, and performance metrics leave you guessing during outages. - **Who Owns This When It Drifts?** — [Source]() — Published 2026-01-28 Tag: Maintenance Model Evaluation suites need ongoing attention and clear ownership. Models drift as real-world data deviates from training distributions. Set-it-and-forget-it reliability doesn't exist in production. You're committing to continuous maintenance whether you planned for it or not. ### In Dialogue With Complexity - [An Interview With the Context Window on Why It Was Never Meant to Be a Security Boundary]() — Published 2026-01-28 We're sitting in what feels like an infinitely expanding room, which is fitting given who we're talking to. The Context Window doesn't have a physical form, but if it did, it would probably look like someone who hasn't slept in months. This is the component at the heart of every large language model, the memory buffer where all input gets processed together. And lately, it's been asked to do something it was never designed for: serve as a security boundary. ## Other sections in Issue 14 - [Daily Brief]() - [Market Pulse]() - [Practitioner's Corner]() - [Echoes]() - [Vision]() ## More - [Publication guide]() - [Issue archive]()