---
title: "Foundations"
canonical: "https://current.tinyfish.ai/issue/55/foundations"
publication: "TinyFish"
publisher: "TinyFish"
language: "en-US"
guide: "https://current.tinyfish.ai/llms.txt"
permalink: "https://current.tinyfish.ai/issue/55/foundations"
issue: 55
section: "foundations"
description: "Conceptual clarity earned from building at scale"
authors: ["Sevda Polat","Sable Whitford"]
published: 2026-11-16
updated: 2026-11-13
articles: 29
---

# 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](<https://current.tinyfish.ai/issue/55/foundations>)

## Issue 55 — Foundations

Conceptual clarity earned from building at scale

### Identity and Access

- [An Agent's Identity Settles the Easiest of Three Questions](<https://current.tinyfish.ai/issue/55/foundations/article/65059/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Core Concepts

You ask an AI assistant to pay one invoice. A prompt appears, you tap approve, and it feels like you agreed to exactly that. But what did the tap actually authorize? A draft standard for governing AI agents offers an odd way into the question. It says it doesn't need to know whether the thing making the request is an agent. For a rule written about agents, that sounds like a gap. Follow it, though, and it leads back to your approve button.

### Failure Walkthrough

- [A Timed-Out Payment Is Unknown, Not Failed](<https://current.tinyfish.ai/issue/55/foundations/article/65079/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · Published 2026-11-13 · Format: Failure Walkthrough

An AI agent asks Stripe to charge a customer's card, and Stripe does it: the card is charged, and Stripe sends back the ID of the new payment. But the connection drops before that reply reaches the agent. The agent waits, raises an error, and its code files the order as declined. The customer has paid, yet the agent's records say the charge never happened. What follows is a walkthrough of how that happens and why your own records end up disagreeing with Stripe's.

### Read the Limitations Section

- **Mapping prompt injections hidden in web pages** — [Source](<https://scnps.co/papers/ccs2026_prompt_injection.pdf>) — Published 2026-11-13

The crawl misses logged-in pages and agent-only content, yet found 15.3K injections in 1.2B URLs.

- **A debugging tool that never timed developers** — [Source](<https://aclanthology.org/2026.emnlp-demos.52.pdf>) — Published 2026-11-13

The authors never measured developer time; the tool recovered 13 of 73 benchmark failures.

- **AI judges scored by a flawed answer key** — [Source](<https://aclanthology.org/2026.emnlp-main.659.pdf>) — Published 2026-11-13

Its answer key marks down valid extra tool-call arguments, and one annotator checked each record.

- **Retrieval benchmark too small for close calls** — [Source](<https://aclanthology.org/2026.aacl-long.245.pdf>) — Published 2026-11-13

With about 60 questions per category, the authors warn, small gaps between models stay unsettled.

- **Routing tasks when an agent abstains** — [Source](<https://aclanthology.org/2026.realm-1.57.pdf>) — Published 2026-11-13

Evidence comes from one e-commerce assistant, where task success rose from 71.5% to 84%.

- **Surveying deployed agents, with a sampling caveat** — [Source](<https://proceedings.mlr.press/v306/pan26a.html>) — Published 2026-11-13

The authors flag sampling bias in their 86 deployed systems, where reliability topped reported problems.

### Debugging Agents by Symptom

Traditional software at least has the decency to crash. An AI agent, meaning a model that plans its own steps and calls tools, browsers, and APIs to carry them out, usually fails politely. The run finishes, the logs look clean, and the agent reports success while the customer is still waiting.

That makes agents miserable to debug. "Done" is built from smaller claims: a query executed, a tool returned something, a click registered. Each can be true while the outcome is wrong.

So the real work is finding which layer made a narrower claim than it seemed to. Gather evidence before you settle on a theory, because one symptom can have several causes.

- **Agent says done, customer got nothing** — [Source](<https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/mcp.md>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Completion check

Start at the destination system and work backward to the agent's summary. Compare what the agent claimed, what the tool returned, and what the target system recorded. A tool call over MCP, a common agent-to-tool protocol, can finish cleanly and still carry `isError: true`. If a write may have landed before a timeout, see this issue's payment walkthrough.

- **Same input, different result on every rerun** — [Source](<https://aclanthology.org/2025.eval4nlp-1.12.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Full input

The user's message is a small slice of what the model actually received. Diff the instructions, model settings, retrieved documents, tool definitions, and memory between runs. One study saw five hosted models vary across ten runs at supposedly deterministic settings. Another saw rankings move when only the date in the system prompt changed.

- **The answer's in our docs. Still wrong.** — [Source](<https://papers.neurips.cc/paper_files/paper/2024/file/27245589131d17368cccdfa990cbf16e-Paper-Datasets_and_Benchmarks_Track.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Retrieval path

Split it into two questions. Did the fact you needed reach the model's context at all? If it did, did the answer actually use it? RAGChecker, an evaluation framework for retrieval systems, scores those separately. It also found that pulling in more context can just hand the model more noise to trip over.

- **A one-time exception became a permanent rule** — [Source](<https://aclanthology.org/2026.findings-emnlp.571.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Agent memory

Follow the memory from start to finish: the original conversation, what got saved, what type it was filed as, and where it came back. The MemGuard paper documents passing events hardening into durable "facts," and anecdotes overriding explicit rules, when stored facts, events, and rules get mixed together.

- **It sees the button but can't click** — [Source](<https://www.w3.org/TR/2026/WD-webdriver2-20261109/>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Browser actions

Save the exact browser error, the frame, the element reference, and a screenshot. The WebDriver automation standard separates three failures: the element went stale after the page changed, something else is covering it, or it can't be interacted with where it sits. A delivered click also doesn't prove it hit the right button.

- **Works for the developer, fails for customers** — [Source](<https://www.rfc-editor.org/rfc/rfc6750.html#section-3.1>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Auth scope

Check which credential each run actually used, since a developer's own token may carry permissions a customer's doesn't. OAuth separates an invalid token (expired, revoked, malformed: HTTP 401) from insufficient scope (HTTP 403). Under MCP, missing scope can hide a tool entirely, so the agent never even attempts the call.

- **It keeps doing the same thing, forever** — [Source](<https://proceedings.nips.cc/paper_files/paper/2025/file/b1041e52d3be19f0a9bc491657488e4a-Paper-Datasets_and_Benchmarks_Track.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Orchestration state

Line up each action next to what the agent observed, then check what state each step could actually see. The MAST study of multi-agent failures separates repeated steps, lost history, missed stopping points, and information one agent never passed on. In one trace, an unshared username format caused endless failed logins.

- **It stops halfway through, with no crash** — [Source](<https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/client-inference.md>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Stop conditions

Grab the model provider's finish reason (`stop`, `length`, `error`), the configured output limit, and the partial response. Text cut off at `length` can read a lot like a finished answer. Then check the agent runtime's own record of why it decided the task was over.

- **The query runs. The answer is wrong.** — [Source](<https://aclanthology.org/2026.emnlp-industry.42.pdf>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Business meaning

Running is a low bar. On FLINT's production financial datasets, seven text-to-SQL methods wrote queries that executed over 99% of the time, and none topped 50% accuracy. Check which business definition the agent pulled, which reference values it resolved, and whether it applied them correctly.

### Which Agent Specs Are Final

If an agent in your stack calls tools, signs into systems, or acts on a website, you're building on spec documents. Some are contracts you can ship against. Others are proposals with a repo and good intentions. In a link preview, the two look identical.

You learn the difference in production. A field marked *experimental* can be ignored by the agent host you run on, and you find out when something runs that shouldn't have, or refuses when it should.

Four pieces of agent plumbing are moving this fall. They sit at very different distances from safe-to-depend-on, and each has one concrete signal worth watching.

- **Revision 2026-07-28 is published and current** — [Source](<https://modelcontextprotocol.io/specification/2026-07-28/server/tools>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: MCP tools

MCP (Model Context Protocol) is how many agents call external tools. Its July 28 release is the current revision, and tools are a section of it. If a tool declares an output schema, servers **must** return conforming results and clients **should** validate them. *Current* still allows compatible changes.

- **Watch: which revision your client actually negotiates** — [Source](<https://github.com/modelcontextprotocol/modelcontextprotocol/blob/main/docs/docs/2026-07-28/learn/versioning.mdx>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: MCP tools

Publication doesn't tell you what your client speaks. Requests declare a protocol version, servers report the versions they support, and both sides can run several at once. Check what you're actually negotiating. Tasks also moved out of core into an extension with its own status.

- **Community draft, with browser origin trials running** — [Source](<https://webmachinelearning.github.io/webmcp/>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: WebMCP draft

WebMCP lets a web page expose JavaScript functions as tools an agent can call. It's a Draft Community Group Report, not a W3C Standard. Chrome 149 and Edge 150 list origin trials (time-limited opt-in tests for sites), plus a local-dev flag. Experiment, and expect the spec to move.

- **Watch: who validates output schemas, and how strictly** — [Source](<https://github.com/webmachinelearning/webmcp/issues/92>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: WebMCP draft

Output schemas are optional today, and agents may reject results that don't match. Issue #92 records a March resolution to put validation in the browser, but the supported JSON Schema subset is undefined and contributors disagree on making schemas mandatory. Watch for actual spec text.

- **Revision 02: an individual draft without IETF standing** — [Source](<https://datatracker.ietf.org/doc/draft-rosenberg-oauth-aauth/>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: AAuth draft

AAuth is a proposed protocol for authorizing AI agents. Revision 02, by Jonathan Rosenberg and Pat White, was posted November 1, 2026. Datatracker lists it as an active individual Internet-Draft, which the IETF does not endorse. The Standards Track label in its header is a goal, not a status.

- **Watch: renewal before the May 5, 2027 expiry** — [Source](<https://datatracker.ietf.org/doc/draft-rosenberg-oauth-aauth/history/>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: AAuth draft

Internet-Drafts expire after six months unless replaced. Revision 01 expired April 22, and 02 didn't appear until November, so gaps happen. Revision 02 expires May 5, 2027. Watch the history for renewal or lapse, and the status record for any formal change, such as working-group adoption.

- **allowed-tools is optional, experimental, and unevenly supported** — [Source](<https://agentskills.io/specification>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Agent Skills

A skill is a folder of instructions and metadata, sometimes with scripts, that an agent loads when relevant. The optional `allowed-tools` field lists tools pre-approved to run, as a space-separated string. The spec marks it experimental and says support varies by agent. Confirm your host honors it.

- **Watch: string or list, or outright replacement** — [Source](<https://github.com/agentskills/agentskills/issues/144>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) — Published 2026-11-13

Tag: Agent Skills

Issue #144 debates whether the field should stay a string or become a YAML list. One contributor quotes project discussion suggesting the field, inherited from the pre-standard format, could give way to a vendor-neutral mechanism. Nothing is adopted. For now, your host's own docs decide what actually happens.

### Past Articles

- [Test a Page's Agent Tools the Way You Test Its Buttons](<https://current.tinyfish.ai/issue/55/foundations/article/61611/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · 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?

- [Revenue Is Number 28, and Someone Has to Tell the SQL Agent](<https://current.tinyfish.ai/issue/55/foundations/article/61610/markdown>) — By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>) · 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.

- [Three Ways an Agent Can See a Page, and What Each One Drops](<https://current.tinyfish.ai/issue/55/foundations/article/57538/markdown>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) · 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](<https://current.tinyfish.ai/issue/55/foundations/article/57518/markdown>) — By [Sable Whitford](<https://current.tinyfish.ai/team/201/markdown>) · 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.

## Other sections in Issue 55

- [Daily Brief](<https://current.tinyfish.ai/issue/55/daily-brief/markdown>)
- [Market Pulse](<https://current.tinyfish.ai/issue/55/market-pulse/markdown>)
- [Practitioner's Corner](<https://current.tinyfish.ai/issue/55/practitioners-corner/markdown>)
- [Echoes](<https://current.tinyfish.ai/issue/55/echoes/markdown>)
- [Vision](<https://current.tinyfish.ai/issue/55/vision/markdown>)

## More

- [Publication guide](<https://current.tinyfish.ai/llms.txt>)
- [Issue archive](<https://current.tinyfish.ai/issues/markdown>)
