--- title: "Foundations" canonical: "https://current.tinyfish.ai/issue/54/foundations" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/54/foundations" issue: 54 section: "foundations" description: "Conceptual clarity earned from building at scale" authors: ["Sevda Polat","Sable Whitford"] published: 2026-11-09 updated: 2026-11-05 articles: 27 --- # 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 54 — Foundations Conceptual clarity earned from building at scale ### Concept Feature - [Revenue Is Number 28, and Someone Has to Tell the SQL Agent]() — By [Sevda Polat]() · Published 2026-11-05 · Format: Core Concepts Ask an AI agent a question about your company's data, and it writes a database query, runs it, and returns a number. New research tested how often that number is right. Running was rarely the problem: every system tested wrote queries that ran at least 99% of the time. Yet apart from the researchers' own system, none answered even half the questions correctly. The wrong answers look as exact as the right ones, and nothing warns you which is which. This close read asks what a query that runs actually tells you. ### Practical Walkthrough - [Test a Page's Agent Tools the Way You Test Its Buttons]() — By [Sevda Polat]() · Published 2026-11-05 · Format: Practical Walkthrough Web pages are starting to hand AI agents tools instead of buttons. Rather than clicking, the agent calls a tool, sends it some inputs, and gets something back. That raises a question clicking never did: who checks what goes in and what comes out? ### Documented Cases - **Clarkesworld Closes Submissions to All Writers** — [Source]() — Published 2026-11-05 Swamped by machine-written stories, the editor closed submissions, locking out human writers too. - **Eighteen Million Fake Net-Neutrality Comments Traced** — [Source]() — Published 2026-11-05 New York's settlements put consent and recordkeeping duties on lead generators who impersonated people. - **CFPB Adds Verification to Complaint Filing** — [Source]() — Published 2026-11-05 After complaints doubled, two-factor checks and third-party labels fall on complainants and their representatives. - **curl Moves Minor Vulnerabilities Into Public** — [Source]() — Published 2026-11-05 To cope with report volume, curl handles low-severity flaws publicly, exposing users earlier. - **One Researcher's AI Reports Fixed curl** — [Source]() — Published 2026-11-05 Amid the junk reports, one researcher's AI-assisted findings produced roughly 50 merged curl fixes. - **AI-Drafted Appeal Overturns a Hospital Bill** — [Source]() — Published 2026-11-05 A family's AI-assisted appeal, plus persistence, got a roughly $1,800 nursery charge dismissed. ### Specification Status Guide An agent that books travel or files a ticket usually depends on documents other people are still writing: a tool protocol, a browser-automation spec, a proposed HTTP header. Each carries a status label, and it is tempting to read those labels as one shared scale of readiness. They come from different institutions with different rules. A W3C community group report marked "Final" still sits outside the standards track. MCP's "Current" revision can still change. An IETF "Proposed Standard" is already a published, approved RFC. Before you schedule a migration, check what each label actually commits to, and which people answer for the text when it moves. - **WebMCP: a community draft outside the standards track** — [Source]() — By [Sevda Polat]() — Published 2026-11-05 Tag: W3C Community The Web Machine Learning Community Group published this Draft Community Group Report on November 2. Participants made patent commitments for their own contributions, and the draft collects implementer feedback. It has had no formal W3C review or endorsement, cannot be cited as a standard, and may never enter the track. - **WebDriver's July draft: standards track, still provisional** — [Source]() — By [Sevda Polat]() — Published 2026-11-05 Tag: W3C Working-Group The Browser Testing and Tools Working Group published this Working Draft on July 2, edited by Simon Stewart and David Burns, under W3C's Process and Patent Policy. The draft disclaims endorsement, may be replaced at any time, and asks to be cited as work in progress. The 2018 WebDriver Recommendation is a separate, finished document. - **MCP's July revision: Current, and still changing** — [Source]() — By [Sevda Polat]() — Published 2026-11-05 Tag: MCP Core The 2026-07-28 release is labelled Current: ready to use, though backward-compatible changes can still land under the same date. Final is reserved for earlier revisions, and implementations may support several at once. MCP belongs to the Agentic AI Foundation (AAIF); Core Maintainers set direction and Lead Maintainers make final calls, all of them individuals. - **Tasks extension: Beta, with a path toward core** — [Source]() — By [Sevda Polat]() — Published 2026-11-05 Tag: MCP Extension Reaching Beta required an Extensions Track proposal and a reference implementation in an official SDK. Breaking changes should be rare and documented, but they are allowed. MCP's lifecycle rules treat a "to Core" designation as a signal of intent, and extensions stay optional for conformance. Older "Stable" labels described schema snapshots. - **Agent Skills: new governance, release process still pending** — [Source]() — By [Sevda Polat]() — Published 2026-11-05 Tag: AAIF Project Anthropic contributed the format to AAIF, which admitted it on October 22. Rules adopted November 4 set up a steering committee of individual maintainers led by Mahesh Murag and David Soria Parra. Numbered proposals and tagged releases are planned but not yet running. Shared stewardship of the format won't make every agent execute skills the same way. - **Idempotency-Key: a working-group draft, not yet an RFC** — [Source]() — By [Sevda Polat]() — Published 2026-11-05 Tag: IETF Draft Revision -08, dated October 22, belongs to the IETF's HTTPAPI working group, so the group controls changes to the text, not a single author. Any Internet-Draft is valid for six months at most, can be replaced at any time, and needs rough consensus to advance. The suffix counts revisions and says nothing about maturity. - **RFC 9396: published and approved, on the first rung** — [Source]() — By [Sevda Polat]() — Published 2026-11-05 Tag: IETF RFC Rich Authorization Requests came out of the OAuth working group in May 2023 with IETF consensus and approval from the IETF's steering group, the IESG. Proposed Standard means a stable specification fit to deploy. It is the entry level, below Internet Standard, needs no prior operational experience, and doesn't tell you how widely it is deployed. ### Agent Strategy Overview Before an agent clicks anything, a lot has already been decided on its behalf: which model gets the request, whether it writes a plan first or works it out as it goes, how the job gets split, how many alternatives stay alive, what a retry learns from, and when it quits digging. Most of those choices live in framework defaults nobody revisits. Each one carries a bet about your task. Maybe it's that you can rewind the environment, or that something will flag a wrong answer, or that you know what evidence the answer needs before you start looking. The papers below are good work, and their results come with those bets attached. Check the bets against your actual production traffic before you borrow the technique. - **Routing assumes your traffic resembles its training data** — [Source]() — By [Sable Whitford]() — Published 2026-11-05 Tag: Model Routing RouteLLM learns from preference comparisons to predict when a cheaper model's answer will hold up, then applies a cost threshold you choose. It presumes labeled comparisons and a settled trade-off. It routes single responses rather than multi-step agent runs, and it weakens as your traffic drifts from its training data. - **Let each observation decide the next move** — [Source]() — By [Sable Whitford]() — Published 2026-11-05 Tag: Interleaved Reasoning ReAct interleaves written reasoning, tool calls, and observations, so one search result can redirect the next query. It assumes someone built an action interface for the task and that observations carry useful signal. The paper itself documents how its deliberately narrow Wikipedia tools limited what the agent could find. - **A trained planner writes the map first** — [Source]() — By [Sable Whitford]() — Published 2026-11-05 Tag: Global Planning EAGLET trains a separate planner to produce a global plan from the instruction, while the executor still acts on its live history. Training needs completion rewards and the ability to run tasks with and without a plan. The authors flag a trap: strong executors succeed despite bad plans, so success alone says little about plan quality. - **Map dependencies first, then run calls in parallel** — [Source]() — By [Sable Whitford]() — Published 2026-11-05 Tag: Task Decomposition LLMCompiler plans tool calls as a dependency graph and fires independent ones concurrently. On HotpotQA questions comparing two entities, it roughly matched a tuned ReAct baseline's accuracy at about half the latency: 3.95 seconds versus 7.12. It assumes tool specs are supplied and dependencies are knowable before execution starts. - **Tree search needs an environment you can rewind** — [Source]() — By [Sable Whitford]() — Published 2026-11-05 Tag: Tree Search Language Agent Tree Search runs Monte Carlo Tree Search over agent trajectories, scored by model value estimates, environment feedback, and reflections. The authors state the prerequisite outright: the agent must revert to earlier states. A live website where a submitted form stays submitted doesn't allow that. They also acknowledge heavier compute costs. - **Retries learn only what the evaluator reports** — [Source]() — By [Sable Whitford]() — Published 2026-11-05 Tag: Reflection Loops Reflexion turns feedback on a failed attempt into written notes for the next try, with no weight updates. It assumes a retryable task and something that scores it. With self-generated tests, results split against the baseline: 91.0% versus 80.1% on HumanEval, but 77.1% versus 80.1% on MBPP, where faulty tests passed faulty code. - **Triggering reflection when there's no answer key** — [Source]() — By [Sable Whitford]() — Published 2026-11-05 Tag: Reflection Triggers REGAR samples several reasoning trajectories, reflects when they disagree, and stops once its beliefs stop shifting between rounds. It assumes candidates can be sampled and compared against preset thresholds. The authors concede that agreement measures consensus rather than correctness, so candidates that are consistently wrong can settle into a stable, confident mistake. - **Stop searching when every required claim has evidence** — [Source]() — By [Sable Whitford]() — Published 2026-11-05 Tag: Stopping Search HALT adds a stopping gate to a retrieval agent: search ends once gathered evidence covers every claim the reasoning is expected to need. That assumes the question breaks into claims before searching begins. Model-generated claims saved fewer searches than benchmark annotations did, and verification raised token processing even as wall-clock time fell. ### Past Articles - [Three Ways an Agent Can See a Page, and What Each One Drops]() — By [Sable Whitford]() · Published 2026-10-30 · Format: Hands-On Walkthrough Before a web agent clicks anything, somebody has already decided which version of the page it gets. The usual choices are a screenshot, the accessibility tree browsers build for screen readers, or the DOM, the browser's full model of the page. Each one makes some questions easy and others impossible. - [Your Agent's Questions Need a Recipient and a Timer]() — By [Sable Whitford]() · Published 2026-10-30 · Format: Core Concepts 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. - [Compared With What? Checking Whether Agent Memory and Skill Libraries Earn Their Tokens]() — By [Sable Whitford]() · Published 2026-10-23 · Format: Core Concepts Agent frameworks keep bolting on memory, self-reflection, and libraries of saved skills, and the benchmark numbers usually go up. Before you credit the add-on, ask what the plain agent was allowed to spend. When researchers gave a basic web agent a few extra steps instead, it kept pace with the enhanced versions and cost less to run. A cancer-records study found the same skill library worked about as well with the agent's planning and tool-calling stripped away. Then there are tasks spanning several sessions, where memory should finally earn its keep. - [The Other Agent Got the Message. Did It Change Anything?]() — By [Sable Whitford]() · Published 2026-10-23 · Format: Practical Walkthrough Wire AI agents together and every hand-off lands in a transcript, which people tend to read as proof the messages mattered. It's a delivery receipt. To find out whether Agent B actually used what Agent A told it, change one fact in one message and see whether B's next move shifts. One study ran that test 793 times. Depending on where the noise bar sat, 39% to 75% of tests found no message driving the decision. And for the notes agents share, saving everything did worse than saving nothing. ## Other sections in Issue 54 - [Daily Brief]() - [Market Pulse]() - [Practitioner's Corner]() - [Echoes]() - [Vision]() ## More - [Publication guide]() - [Issue archive]()