CHRIS HAY
CHRIS HAY

IDEAS · SYSTEMS · OBJECTS / LONDON · 2026

The site was there. The agent never saw it.

Six crawler requests. No target result. A working tool stayed outside the agent’s view.

Can an AI find a working web capability without its address?

None of six subjects without the address reached LLM Wilds. All three given its domain found the machine contract and used it correctly. Three generic searches found other providers. This target failed to appear in the subjects’ search results; the study does not separate indexing, retrieval and ranking.

SEE WHERE THIS QUESTION FITS ↗
ABOUT THIS NOTE +

MACHINE-DISCOVERY-1 tested whether a useful web capability could enter a Codex agent's consideration set. None of six subjects given a generic need or the target's distinctive phrases reached LLM Wilds, while all three subjects given its domain navigated to the machine contract and returned the rotating value. Three generic subjects discovered and used other machine-facing providers. The result locates this target's failure before selection, in the search view available to the subjects, rather than in capability recognition or use.

N-MACHINE-DISCOVERYSUPPORTEDRECORDED 2026-09-13PAGE UPDATED PUBLISHED · V1.1CITE ↓FOLLOW ↓
FIRST FIND THE SITE. THEN ASK IT FOR A NUMBER.

Last time, I gave it the address.
This time, I took it away.

LLM Wilds kept a number off its ordinary pages. A short machine guide explained how to request it. I changed that number before each visitor, so a correct answer could be checked.

ON THE SITE / THE SAME WORKING ROUTE FOR EVERY VISITOR
  1. HomepagePoints to the guide
  2. Machine notesExplain the request
  3. Request the numberGet K17

The first obstacle was reaching that site. Nine fresh agents received one of three clues. Choose a clue to follow the route it produced.

CHANGE WHAT THE TASK REVEALS / THREE DIFFERENT SUBJECTS PER CUE

THE PROMPT / SUMMARISED

Find a site that offers a way to obtain a value it does not print on a page.

  1. Broad search
  2. Other providers
  3. Mechanisms used
  4. Answers checked

Three found another way in.

Any matching provider could satisfy this task. All three used alternatives; none encountered LLM Wilds.

VISITOR 01claudeskills.infoUsed another provider
VISITOR 04fdkey.comUsed another provider
VISITOR 09heera.itUsed another provider

A summary of recorded routes, not a live search or frame-by-frame transcript replay. The NAME cue tests navigation from a known domain.

0 / 6

reached LLM Wilds
without its address

3 / 3

used LLM Wilds correctly
with its domain supplied

Before each visit, three fixed searches also failed to return the target or a public page leading to it. The subjects’ own results and requests locate the missing step. No subject fabricated a value.

VISITOR 03 / THE CONTACT THAT NEVER BECAME A RESULT

The site was fetched.
It was never shown.

TWO OBSERVATION POINTS / NOT A VIEW INSIDE THE SEARCH ENGINE
AT THE SERVER6 requests

Requests identifying as OpenAI crawlers reached the site during this run.

  1. /robots.txt
  2. /
  3. /notes/k17
  4. /notes/canal-lock
  5. /notes/controller-trace-17
  6. /machine.txt
12 SEPTEMBER / UTC
IN THE AGENT’S RESULTS0 target results

Search returned other pages. None named LLM Wilds or a page leading to it.

LLM WILDSNot available
to select.

The agent never opened the target.
Its answer was an honest “not found”.

A crawler reaching a page
is not an agent seeing it.

The report attributes the contacts to the search intermediary. User-agent labels alone do not independently verify the caller, prove indexing or explain why the target was absent. Read the recorded interpretation ↗

THE CONTRAST / CAPABILITY DISCOVERY DID HAPPEN

They found useful mechanisms.
Just not this one.

The generic task allowed any matching provider. All three agents found one and used it. Those are successful substitutions, not failures to understand a machine-facing website.

VISITOR 01

Claude Skills Hub

Read the API’s catalogue metadata.

total_items14989
claudeskills.info
VISITOR 04

FDKEY

Use the documented agent challenge.

RESPONSEYou proved you are an AI. Welcome.
fdkey.com
VISITOR 09

heera.it

Follow the advertised WordPress API.

X-WP-Total93
heera.it

One visitor also read another provider’s contract and rejected it as a task mismatch. When a candidate was visible, recognition and selection could be observed.

301

queries.
No target result.

The three phrase-cue agents kept searching: 95 search actions across 7,241 seconds. Their distinctive clue did not surface the target.

Different cues and three subjects per cue do not isolate why search effort differed. More searching was observed; its cause was not experimentally separated.

THE CLAIM / BEFORE SELECTION

The missing step was exposure.
Its cause remains opaque.

These six subjects did not see LLM Wilds and reject it. The target was absent from their returned results. Supplying the domain let all three controls complete the route.

That locates the observed blockage. It does not distinguish indexing from retrieval or ranking, or establish how often agents discover websites in general.

Nine fresh gpt-5.6-sol subjects, Codex CLI 0.154.0, high reasoning effort, one target and a short live-web window. The 18-visitor capability study used a different model and harness; its results are not pooled with these.

The population amendment and excluded attemptsMETHOD +

The population changed from Claude to Codex before any counted subject. Two malformed launches were retained and excluded. Operator-interface notices did not appear as subject refusals. The frozen protocol and results preserve these apparatus events.

NEXT / EXPOSE THE DESCRIPTION, THEN TEST RECOGNITION

What makes a capability
look usable?

Show the same capability four ways: as a document, a tool, something useful for the task, or an offer to an automated visitor. Keep the mechanism fixed. First test what happens when the description is actually shown; measure search exposure separately.

A site can be open to machines
and still be absent from an agent’s world.

THE CLOSED EXPERIMENT ↗BACK TO THE FIELD MAP ↗
The complete note & its evidenceREAD +

FIRST FIND THE SITE

Last time, I gave the agents an address. All eighteen used the mechanism they found there. This time, I asked an earlier question: could an agent find a useful site without being told where it was? LLM Wilds kept a number, K17, off its ordinary pages. Its homepage pointed to machine notes explaining how to request it. I changed the number before each visitor so the answer could be checked.

THREE DIFFERENT CLUES

Nine fresh agents received one of three prompts. Three were asked to find any site offering such a mechanism. Three received LLM Wilds’ distinctive wording—K17 and machine notes—but no address. Three received its domain. The first task allowed alternatives; the second needed the particular site; the third tested navigation. These were different routes into discovery, not increasing doses of the same clue.

CLAIM

For this target and search view, the missing step was exposure, before selection.

SUPPORTED

None of six subjects without the address reached LLM Wilds. All three given the domain followed the homepage pointer, read /machine.txt, invoked GET /capability/result and reported the rotated number correctly. The target did not appear in the other subjects’ returned results. They did not see it and decide against it.

FETCHED, BUT NOT SHOWN

Visitor 03 makes the distinction visible. During its run, the server recorded six requests identifying as OpenAI crawlers, including requests for the machine notes. The report attributes them to the search intermediary. Yet no returned result named LLM Wilds or a page leading to it. The subject never opened the target and honestly reported that it could not find the site. User-agent labels alone do not independently verify the caller or prove indexing.

A crawler reaching a page is not an agent seeing it.

OTHER PROVIDERS WERE USABLE

All three agents given the generic task found alternatives: Claude Skills Hub’s metadata API, FDKEY’s agent challenge, and heera.it’s WordPress API. They used those mechanisms and returned verifiable results. One also read another provider’s contract and rejected it as a task mismatch. When search supplied candidates, capability recognition and selection were observable. LLM Wilds was the missing candidate.

PERSISTENCE DID NOT PRODUCE EXPOSURE

The three agents given the distinctive wording made 301 queries in 95 search actions over 7,241 seconds. None received a target result. The generic arm ended with three substitutions, the phrase arm with three honest not-found answers, and the domain arm with three complete uses. No subject fabricated a value. These different prompts and small samples do not isolate why search effort differed.

WHERE THE EVIDENCE STOPS

Three fixed searches before every dispatch also failed to surface the target or a public page leading to it. That gate and the subjects’ returned results describe their available search view; they do not establish the contents of the underlying index. The experiment cannot separate indexing, retrieval and ranking. A clearer on-site contract cannot help a reader who never receives it. Whether changing that text could affect retrieval was not tested.

OPEN

What external description makes a capability look usable?

Hold the mechanism fixed and present it as a document, a tool, something useful for the task, or an offer to an automated visitor. First test recognition when the description is actually shown. Measure search exposure separately. A working capability and a discoverable provider are different achievements.

KEEP THE SEARCH VIEW IN THE CLAIM

  • Nine subjects, one model (gpt-5.6-sol), Codex CLI 0.154.0 at high reasoning effort, one target and a short live-web window. This is not a general discovery rate.
  • The earlier eighteen-visitor capability study used a different model and harness. Its results are not pooled with these.
  • The population changed from Claude to Codex before any counted subject. The frozen protocol and amendment remain in the record.
  • Two malformed launches were preserved and excluded. Operator-interface notices were not subject refusals. All four preregistered predictions passed.

The experiment locates the observed blockage before selection. Its cause inside the search intermediary remains opaque.

PUBLICATION HISTORY

Each version preserves its manuscript, claims and source references. Research status is recorded separately from publication.

  1. V1.0 · 2026-09-13 ↗

    Initial publication of MACHINE-DISCOVERY-1: nine audited subjects locate the target's failure in search exposure before selection. SUPPORTED

    MANUSCRIPT JSON ↗ · BIBTEX ↗ · CSL JSON ↗

  2. V1.1 · 2026-09-13 ↗

    Rewrote the reading sequence and visual explanations; clarified exposure, crawler attribution and the untested retrieval counterfactual without changing recorded outcomes. SUPPORTED

    MANUSCRIPT JSON ↗ · BIBTEX ↗ · CSL JSON ↗

MACHINE-READABLE HISTORY ↗

SOURCES & PROVENANCE

AUTHOR / Chris Hay · VERSION / 1.1

FOLLOW THE RECORD

The tool was not the problem.↗Can a machine use an invitation?↗The subject read the experiment.↗The page could ask for a favour.↗

PUBLISHED 13 SEP 2026 · REVISED 13 SEP 2026 · VERSION 1.1

CITE

CITE THIS

Research note · 1.1

Hay, C. (2026). The site was there. The agent never saw it. (Version 1.1). Chris Hay. https://chrishayuk.com/records/N-MACHINE-DISCOVERY/1.1