AI Agent Readiness: An Engineer’s Field Notes

An independent technical investigation series

AI agents are coming aboard. Has anybody checked the gangway?

Websites are being asked to do more than load quickly and look respectable. They increasingly need to expose reliable information to software that can retrieve it, interpret it and, in some circumstances, act on it.

Cloudflare and other technology providers are introducing tools for machine-readable content, content-use signals, API discovery, agent interfaces and access controls. Some may prove genuinely useful. Others may add complexity without solving a business problem. The trick is knowing which is which.

At Sydney Business Web, we'll investigate these technologies on our own working infrastructure before recommending them to clients. Each field note will explain the purpose, the configuration, what our tests actually showed, what remains uncertain and whether a business has a good reason to implement it.


Our working method

Examine. Analyse. Act — or don't. Write.

We're not here to switch on every beta feature or chase a perfect readiness score. We're here to establish what works, under what conditions, at what cost and with which risks.

01 / EXAMINE

Understand the claim

Read the documentation, identify the mechanism and define what a successful test would show.

02 / ANALYSE

Gather evidence

Measure actual HTTP responses, content, identity relationships, failures and side effects.

03 / DECIDE

Act — or don't act

Deploy selectively, mitigate risks or leave a feature disabled when the case is unproven.

04 / WRITE

Publish the findings

Record observations, limits and commercial implications in a permanent reference.


Published investigations link to their full technical reports. Possible future topics remain clearly identified as planned or candidates.

Field notes

Each numbered investigation will become a permanent technical page. We will add links when the corresponding write-up is published, rather than pretending unfinished experiments have conclusions.

Published · 28 October 2026

Cloudflare Markdown for Agents: What Actually Changes?

We enabled Cloudflare's Markdown conversion on one page only — the Sydney Business Web homepage — and requested it using Accept: text/markdown. We inspected response headers, converted content and our connected business-identity JSON-LD.

Verified evidence: HTTP 200; Content-Type: text/markdown; all six JSON-LD objects equal to their original parsed structures, including values and relationships. We have not established token savings, reliable agent uptake or improved AI recommendations.

Read Investigation 01 — Cloudflare Markdown for Agents →

Published · 29 October 2026

Cloudflare Agent Readiness: When Does a Website Need to Become an Agent?

Thirteen readiness checks and three architectural questions examined against a live Sydney Business Web production site. We distinguished machine-readable content, interface discovery and secure agent operation — and declined unnecessary infrastructure.

Verified evidence: all five foundation checks reported complete; a homepage Link Header independently verified in HTML and Markdown responses; the automated Link Header audit aborted rather than returning a conclusive result. A readiness score is not evidence of AI recommendations.

Read Investigation 02 →

Coming soon · Planned investigation

Content Signals: Who May Use Your Website's Information?

An investigation into Cloudflare Content Signals and the policies website operators can communicate to machine systems about AI training, search and content use.

We'll examine the actual HTTP headers, what the signals express, whether they are recognised by intended recipients and what they cannot technically enforce.

Status: Investigation reserved. Testing and findings have not yet been published.

Published · 1 November 2026

AI Agent Readiness: From Discovery to Autonomous Execution

A PageSpeed Agentic Browsing result of 3/4 prompted an investigation into Agentic Resource Discovery (ARD). We published a discovery manifest, documented our AI Baby Gorilla diagnostic through OpenAPI 3.1 and subsequently verified a 4/4 result.

But a valid manifest was only the beginning. We extended our independently developed Baby Gorilla diagnostic with a public agent-facing API running on Cloudflare Workers, supported by Cloudflare D1 for quota management and execution auditing.

Verified evidence: ARD manifest validity confirmed in PageSpeed; successful anonymous HTTP 200 API invocation; invalid Bearer tokens rejected; per-requester quota enforcement tested; and a successful production scan recorded in D1. Full global-limit concurrency testing and independent third-party agent adoption remain unproven.

Read Investigation 04 — Baby Gorilla Public Agent API →

Published · 2 November 2026

PageSpeed Insights Agentic Browsing: What Does a 4/4 Result Actually Prove?

We investigated why Google's Agentic Browsing report moved from 3/4 to 4/4 after an ARD manifest was published while Lighthouse's ARD discovery implementation was also changing. Rather than assume a single cause, we tested Lighthouse 13.5.0 and 13.6.0 against the same production homepage, approximately two minutes apart.

Verified evidence: Lighthouse 13.5.0 reported three applicable checks passed (3/3), with ARD marked Not Applicable and assigned zero weight. Lighthouse 13.6.0 reported four passed (4/4), with ARD valid and assigned weight one. Both recorded layout shift of approximately 0.006. The original historical 3/4 result remains unexplained by these controlled runs alone.

Read Investigation 05 — Lighthouse 13.5 vs 13.6 →

Future topic Candidate

Authentication, Permissions and Safe Agent Actions

How to distinguish useful automation from handing unfamiliar software the keys to the engine room.


What we investigate

Not just whether a feature switches on.

  • Retrieval: Can the intended system actually obtain the information?
  • Integrity: Are business facts, entity identifiers and relationships preserved?
  • Interpretation: Is the content useful and unambiguous for the intended machine reader?
  • Operation: What happens to caching, monitoring, performance and security?
  • Commercial value: Does the capability solve a real problem worth its cost?

What we won't pretend

A successful request isn't a guaranteed recommendation.

A crawler successfully receiving a page is evidence of access — not proof that an AI model understands the business, trusts every claim or will recommend it to a customer.

We also won't imply that experimental standards are universally supported, that every business needs agent interfaces, or that turning on a platform feature automatically delivers a commercial benefit.

Engineering isn't the art of making everything complicated. It's the discipline of knowing what needs to work — and proving that it does.


The bigger picture

Readiness begins with a coherent business identity.

Machine-friendly formats and agent interfaces are only part of the story. A business website still needs accessible content, consistent identity, clear relationships between people and services, suitable controls and corroborating evidence. That's why these field notes sit alongside our work on Schema Gorilla and AI Observatory: one examines connected machine-readable identity; the other measures observable machine retrieval.

Those are different questions. Neither is answered by a colourful dashboard alone.

For related engineering groundwork, see our earlier Agentic Browsing Readiness investigation, AI Crawler Monitoring at the Cloudflare Edge, and our findings on why AI systems sometimes cannot retrieve websites, ModSecurity retrieval failures and HTTP 429 responses. Those are separate operational problems; Markdown conversion cannot fix blocked or inconsistent origin responses.