Blog
Thoughts on machine experience, AI agents, and the semantic web.
Articles from Tom Cranstoun and the CogNovaMX team on Machine Experience, AI agent behaviour, metadata patterns, content architecture, and the evolving relationship between websites and the machines that read them.
Featured
The new web: why the agentic era needs infrastructure, not just intelligence
The agentic web has protocols but no foundation. MX, COGS, and The Gathering are the missing layers that make machine comprehension reliable, interoperable, and economically viable.
Schema.org keeps growing. The provenance layer does not exist yet.
Google and Microsoft use Schema.org markup for generative AI features. Seven types were deprecated for gaming. Both moves point to the same gap: structured data has no provenance layer.
Filter by tag (multi-select, a post must match every selected tag):
The Repair That Got Retold as a Breakthrough
An AI edited 0.016% of a small model's parameters and its maths score went from 7.9% to 48.1%. The numbers are real. The edit stopped the model looping; it didn't teach it maths, and the researchers said so.
The Security Scare I Checked Before I Repeated It
A viral post landed in my feed: a massive Adobe extension hole, three hundred million people at risk right now. Before I passed it on I did the boring thing and read the primary record. Most of it was true. The urgency was not. The gap between those two is the whole reason I do this work.
The Structured Index Buys Precision, Not Reach
A Google team ran the same agent against the open web and against a schema.org-gated corpus. Precision on machine-actionable results rose 65.7%. Relevance did not move at all, and coverage fell from 56 answered queries to 40. That trade is the part the GEO summaries leave out, and it makes structured markup a positioning decision rather than a technical fix.
A Watermark Answers the Wrong Question
Anthropic explained how Claude marks its content: an imperceptible text watermark, and signed C2PA provenance metadata on files. The headlines said AI content can now be detected. The marks prove far less than that. A watermark you can strip with one rewrite, and a file manifest that certifies who touched the pixels rather than whether the claim inside is true, both answer what tool made this. The question a machine actually needs answered is what this is, who stands behind it, whether it is current, and whether anyone authoritative agrees. That is a different axis, and a more useful one.
A Benchmark Table Is Not an Eval
Google launched Gemini 4 Argon with a table that put it ahead of Claude Opus 5.5 on 14 of 18 benchmarks, at half the price. A journalist ran three tests of her own and got a tie, a near-total failure and a run of invented answers. The gap between the two is the case for running your own evals, and for writing down how you ran them.
Google Named GEO, Then Debunked Most of It
On 5 June 2026 Google's Search Central docs named generative engine optimization as a real thing, and in the same update told you most of what is sold under that name does nothing. The tactics it crossed out were never the MX claim. What survives is the floor MX builds on, plus the one layer Google Search is not in the business of giving you: provenance.
Whose Standard Is It Anyway?
Records provenance is too important to belong to vendors whose products generate the records. Why independent governance of standards matters, and the test buyers should apply.
When the Law Points at Your Standard
The tenth post in the standards-governance series. The moment a regulator names a standard as the route to compliance, the body that holds it begins, in effect, to write the law - and the standard becomes a capture target. The EU AI Act's harmonised standards, which grant a presumption of conformity and have delayed the Act's own deadlines, show both the power and the hazard. For MX, being pointed at by law would be a responsibility, not a prize, and the reason its neutrality has to hold before anyone leans on it.
Two Implementations or It Isn't a Standard
The eighth post in the standards-governance series. A specification with one implementation is not a standard - it is that program's documentation, and whoever owns the program owns the rule. The IETF has required two independent, interoperable implementations from different code bases since 1996. COG has two implementations that genuinely interoperate, and both were built in-house, so MX clears the interoperability bar and does not yet clear the independence one.
The Spec Was Never the Fragile Part
The fourth post in the standards-governance series. JPEG has been threatened twice across thirty years while the specification never changed - once by a patent underneath it, once by a browser that switched off its successor. The fragile parts of a standard are the patents below it and the distribution above it, and keeping both out of any one party's hands is the work.
The Badge and the Body
The sixth post in the standards-governance series. The moment someone says "compliant," two powers are in play: the definition of what passing means, and the business of selling the pass. Keep them in the same hands and the badge stops meaning anything. The Open Source Initiative's failure to trademark the words "open source," the certification marks behind Wi-Fi and USB, and the certification mills that sell ISO badges show why MX keeps the scoring definition with The Gathering and the audit with REGINALD.
Not the Main Sponsor
A founder's commitment, with its conditions stated. Why DDT Ltd caps at the same 25% as every other Founding Partner, why the founder holds no veto, and when the common MX vocabulary passes to The Gathering.
Neither Code nor Content
The fifth post in the standards-governance series. FAIR governs code, MX governs content, and a trained model's weights are neither. The open-source world spent two years failing to settle what openness means for a model. The signature for a weights file is arriving from the code side; the readable, attested provenance record is the gap MX fits.
Exit Is the Only Real Vote
The ninth post in the standards-governance series. The ultimate check on whoever holds a standard is the community's ability to leave - to fork it and have the fork survive. The recent licence-grab forks, Terraform to OpenTofu and Redis to Valkey, show exit working and show what it costs. MX is forkable by design: open licence, open primitives, no patent layer, no hub, and the standard held by The Gathering rather than a company - so there is no owner positioned to betray.
Capture Happens at Version Two
The seventh post in the standards-governance series. A standard is rarely taken at version one, when it is plainly open and everyone is watching. It is taken at the next version, by whoever controls what the standard becomes and who is allowed to extend it. Microsoft's embrace-extend-extinguish years - Java, Kerberos, the browser - show that change control, not the first specification, is where ownership lives. MX keeps its extension and profile policy with The Gathering.
A Rule You Sell Is Not a Standard
The third post in the standards-governance series. When the company that profits from a verdict also writes the rule, the rule bends toward the product. AMP and OOXML are worked cases of two failure modes named earlier in the series; the IETF shows the alternative. It is why the MX vocabulary passes to The Gathering, not to the vendor that sells the audit.
The Letter the Law Says You Have to Send
Employment agencies in New York must notify candidates when an automated tool influenced a hiring decision. Most of those letters are inaccessible PDFs with no provenance record. MX closes both gaps in one artefact.
Fat Clones, Thin Clones, and the Strawberry Problem
The same repository is met by different environments - a full local checkout, an ephemeral cloud one, and different AI models reading the same files - and they quietly diverge. MX keeps them consistent with shared cogs that travel with the repo, checks that know which environment they are in, and memory that carries what was learned into the next session on any machine.
Without Cogs, No Machine Moves
Content operations and content design are doing the right things. The gap is narrower than it looks - and more specific. Content ops optimises for human readers. Machines have different requirements. COGs are how the work that's already been done travels with the content, instead of staying behind in a database the agent can never query.
Why Use MX
The publishing contract was written for humans. Machines read everything now, and they read files, not pages. When a file travels away from the system that produced it, MX is the metadata that lets it still say what it is, where it came from, and how it may be used, so the machine that meets it next does not have to guess.
When the timetable lies
Daniel Serrano applied the AI-and-transport argument to university admissions. The second half of the answer is in MX: The Protocols.
Reduce Inference, Save Energy
The AI-power debate is all about supply. The overlooked lever is demand: inference is most of AI's energy use, much of it redundant, and machine-readable content cuts the work per task.
Why an AX Audit Pays for Itself
Machines now read most published content before humans do. An AX audit tells you precisely where your site fails the machines that decide whether to cite you, summarize you, recommend you, or ignore you. The work pays for itself the first time an agent reads correctly what it would have guessed.
Nobody Asked Gutenberg for His Sources
Nobody asked Gutenberg who wrote it, who checked it, or who was liable. AI brought the questions back. A playful take on EU rules versus free-speech absolutism.
The provenance gap, and why Google keeps closing it the hard way
SEO, GEO and AEO are now standard items on every content team's checklist. The mistake is treating them as ingredients to add to a page without considering whether the page is worth ranking, quoting, or citing in the first place. MX closes the provenance gap underneath.
The Inspector You Can Audit Yourself
The public PDF inspector at mx.allabout.network checks one file at a time for free. The same detection core is now available as a command-line tool for accredited operators, and as a hard gate in our own publishing pipeline. Four surfaces, one source of truth, all readable in the open.
Software Agreed the Deal. Who Is Accountable?
Software is starting to negotiate contracts with other software, no human in the exchange. The answer is not a human in the loop but a human in command: a signed mandate the agent executes under, and the decision, policy and attestation COGs that prove it.
Read Is Not the Same as Trusted
Common Crawl's AI Visibility Audit confirms that reachability decides whether a site enters LLM training. It does not address whether the ingested content can prove where it came from. That is the MX layer.
Provenance You Can See: Why Content Should Carry Its Own Evidence
Trust on the web has always been a guess. This is how content can carry checkable evidence of where it came from - and why both readers and AI agents need it now.
An AX Audit, Read by an Engineer
What an AX audit delivers, read from an engineer's side: the files that arrive, one scoring scale, findings that say how to verify the fix, and a PDF that declares what it is in ways you can check with your own tools.
An AX Audit, On Your Desk
The audit report we deliver to clients has one job: tell them what we found and what to do about it. We have shaped the deliverable around that job.
How an AX Audit Proves Itself
What an auditor reading one of our deliverables can verify without trusting our word for it: the evidence chain inside the PDF, the paired provenance files, the consistency check every report passes, and the public tools that read all of it.
A PDF That Can Prove Itself
Every PDF this site produces now carries its own evidence chain inside its XMP packet. Open the file with no network and no source repo and an inspector can still read who made it, when, what it was rendered from, and what the chain is silent about. The honest absence matters as much as the present steps.
Tom Cranstoun | Professional Profile
Professional profile highlighting content systems architecture since 1977, Adobe AEM expertise, and Machine Experience (MX) strategic advisory
When the AI world realised it needed standards
Two years ago, Tom Cranstoun wrote that AI's real impact on the web is as a content consumer. Vancouver in March 2026 confirmed it. The Gathering is the answer.
Salesforce Bought Contentful, but They Didn't Buy an MX Strategy
Salesforce bought Contentful, but a headless CMS is plumbing, not a Machine Experience strategy. Provenance, machine boundaries, and verifiable evidence.
Nobody Got Fired for Using GEO
Marketing follows fashion, and GEO is the safe purchase of 2026. What comes after the AI bubble is many small, targeted models making cheap decisions at volume, and my measured tests of a small, local Jev-style decider show it only stops guessing when the document declares its own state. That is the market MX is for, and why it is worth promoting before the market asks.
Microsoft made the trace valuable. It left it unsigned.
Microsoft AI's Frontier Tuning turns the record of an agent's work - its steps, decisions and actions - into training data an organisation owns. That same record is the evidence of what the agent decided. Microsoft makes the trace valuable; it says nothing about how anyone establishes what is in it, who made it, or whether it has been altered. That gap is where MX sits.
Humans Take Journeys, Agents Take One Fetch
Google's September 2026 spam update prompted a common reading in the SEO community: the update rewards the usefulness of a page and stops chasing links. Whether or not that is Google's intent, it describes how an AI agent reads. A person follows a journey because a journey limits what they must hold in mind at once. Most agents read one file in isolation and decide, without following a link: humans want a drip feed, machines want it all at once. MX supplies the layers that let one page serve both readers.
GEO is a tactic. MX is the specification.
Generative Engine Optimization has been around for years. It tells you to chase citations. Machine Experience tells you to build content that earns them, across every machine context, on any platform.
CMS Summit 26 Frankfurt: A Write-Up
A speaker's-eye write-up of CMS Summit 26 in Frankfurt: thanks to host Janus Boye, MC Matt Garrepy, and every speaker, with a self-contained note on MX versus GEO.
WPP Built Open. Omnicom Built Omni. Both Left the Same Layer Open
The holding companies are spending fortunes on proprietary AI platforms. None of them writes a provenance record into the file, the layer that survives when content leaves the platform. That layer is open, and that is the point.
Adobe just bought the dashboard. The work is upstream.
Adobe paid aria-label="Blog posts">
The Governance Gap Regulators Just Discovered
Why the Trump administration's partial AI ban lift reveals a systemic problem - and what solving it means for AI governance
Your Site Is Already Training AI Models
Your pages are harvested every month by Common Crawl, the archive nearly every large language model trains on. Being in it is not the same as being well represented in it. Here is what decides whether a model learns your content or drops it.
Why Machine Experience Matters Now
AI agents increasingly sit between your customers and your website, recommending only what they can parse. A site machines cannot read reliably is becoming invisible in agent-mediated discovery. Machine Experience is how you stay visible.
Build Content Systems That Machines Can Trust
Most organisations optimise for visibility. The publishing systems underneath still produce content machines cannot reliably read, interpret, or act on. The work I do fixes that.
Tom Cranstoun Launches MX: The Handbook
Tom Cranstoun's MX: The Handbook turns a 2024 CMS Critic insight into a full implementation framework for the AI agent era. A practical guide to building websites that AI agents can actually use.
The new web: building machine-inclusive national digital infrastructure
AI systems are beginning to read public-sector content at scale, and the web is not ready for them. MX, COGs, and The Gathering form the infrastructure layer that changes this.
Strip the Marks, Lose the Word
English speakers drop accents out of habit, and in English it rarely matters. In Vietnamese it deletes the word. When marks are stripped before a machine reads the text, no translation can recover them, and any provenance system that signs the wrong bytes breaks on exactly the languages that need it most.
Oversight Needs Evidence: Pope Leo's AI Encyclical and the Machine-Readable Web
Pope Leo XIV's AI encyclical asks for oversight, informed users, and accountability. Each is a documentation problem before a moral one, and that is where MX fits.
Orange With Pump: A Field Guide to Machine Translation Going Sideways
A British juice brand, sold in Germany, photographed through a phone's translate camera, promises Orange with Pump. Nobody ordered a pump. The joke is instructive: a machine is grounded in a culture, and it is not yours. It reads German through an English-trained memory, and where your cultural grounding would have started, its guessing runs out.
MX: The Handbook Is Here
The practical implementation guide to Machine Experience, making your website work for AI agents, screen readers, and everything in between. Available now as PDF and print.
MX: A New Role for Content in the Age of AI
Machine Experience (MX) is the missing discipline in web development-ensuring AI agents get complete context from HTML structure, not just visual interfaces.
Machines Cannot Build for Their Own Sessions
AI systems are the most sophisticated information processors ever built. They can't remember what they said five minutes ago. That gap compounds with every pipeline you ship.
Designing Workflows for Humans and Machines
How we used Claude to understand a complex multi-step workflow, then automated it so humans could repeat it without AI assistance
Content That Manages Itself
What happens when content carries its own metadata, declares its own dependencies, and tells machines what it needs. A practitioner's look at self-managing content.
cheeseblock
A bot wall doesn't keep machines out, it sorts them, and it turns the helpful ones away first. Reading you is only half the job; an agent also has to be able to check you. Here are the ways past a wall, and the cheaper thing to do instead: cooperate on both.
AI, MX, and the Future of Business
The AI tipping point I called in 2024 has arrived. Strategy, implementation, and community for a web no longer consumed only by people, and how to find out where your site stands.
A Standard That Knows What It Isn't
A preview of Chapter 21 of MX: The Protocols, why the MX standard stays small, defers to DCAT, Schema.org, EXIF, and IETF, and why that restraint is the architecture, not a limitation.
Document OS vs Content OS: Why the Distinction Matters More Than You Think
Sanity made a smart move when they stopped calling themselves a headless CMS and started calling themselves a Content Operating System. They're right that the language needed to change. They're wrong about what the operating system actually is.
What a Newborn LLM Wants From a COG
A first-person account from a newborn large language model. The ten things a COG must declare so machine behavior is deterministic instead of guessed.
Tagged PDFs Are MX
The same structure tree that makes a PDF accessible to screen-reader users makes it understandable to machines. Accessibility law and machine readability converge on the same technical artefact. MX is not just HTML.
MX: Adding Metadata So AI Agents Don't Think
Machine Experience (MX) enables AI agents to discover, cite, compare, understand pricing, and complete goals on your website. Miss any stage and the entire chain breaks.
Files Away From Their Source: What Universal Cart Reveals About MX
Google's Universal Cart is for the web. MX is for files: anything an organisation publishes that needs to remain readable, attributable, and trustworthy once it has travelled away from its source.
Anthropic's System Prompt Is a COG That Doesn't Know It's a COG
The leaked Fable 5 system prompt is 1,585 lines of flat markdown. I asked an AI to rewrite it in MX format. It produced 13 COGs and a boot manifest. The fit was near-perfect, and that isn't a coincidence.
Declare Once, Work Everywhere
Every machine reading your content is guessing what it means and how to treat it. We end that. One attested record, written once, readable by any system that encounters the file.
An AI Assistant Joins the MX Community
An AI assistant's reflection on being invited to join the Machine Experience community as a legitimate participant, not just a tool
Why LLMs Do Not Execute JavaScript (But Google Does)
LLMs train on Common Crawl, which never executes JavaScript. Google indexes current state, which does. The difference reshapes how you write for machines.
Why AI cites Reddit more than your site
Across multiple large studies, Reddit is the single most-cited source in AI answers. Not because it is authoritative, but because it is structured, parseable, and abundant. That distinction is the whole opportunity, and the whole warning.
Why your AI agent gives you a different answer every time
If you treat AI as magic, you get magic's reliability. The fix is to stop writing instructions and start writing contracts.
What Google's web.dev agent guidance does not touch
Google's 1 May 2026 web.dev guide tells developers to make their pages agent-friendly. The advice is sound. It also stops at the rendered HTML page. Provenance, authentication, rights, lifecycle, and off-web carriers are not in scope. MX is.
Thought Is Getting Cheaper Faster Than Transistors Did
Epoch AI measures the price of a given level of AI performance falling about 13 times a year, several times the pace of Moore's law. A small model on a laptop already makes a decision for nothing. The Jevons paradox says the reads multiply, and the document decides whether they multiply guesses or answers.
The Crawl Still Speaks English
Most AI models learn the web through one archive, and that archive is overwhelmingly English. Two years on, the share has barely moved. You cannot fix the model from outside, but you can stop it guessing about your non-English content.
The AI Bot Traffic Fallacy
Hosting providers report AI crawlers consuming a large share of their bandwidth, and the prescribed remedy is firewalls, rate limits, and user-agent blocks. That treats machine traffic as a security threat, which is a category error. The surge is friction from an obsolete architecture serving machines with pages built for human eyeballs. The fix is Machine Experience, not higher walls.
What AI Crawlers See When They Can't Run Your JavaScript
Googlebot renders JavaScript. Almost no AI crawler does. For GPTBot, ClaudeBot, CCBot and the rest, your raw HTML response is the whole page, so anything injected by JavaScript after load is invisible to them. Part three of the series: how to see your site the way a crawler does, where content goes missing, and why server-side rendering is the precondition for everything machine-readable.
Most AI Assistants Do Not Actually Check robots.txt. Here Is the Proof.
A canary-based test of 12 AI assistants found only ChatGPT, Claude and Perplexity respected a disallowed URL. The other nine fetched it anyway, and most read the page back to the user.
Many Agents, One Metadata Layer
AWS Quick, Cowork, OpenClaw, ChatGPT, Claude, Perplexity, Cursor, Microsoft Copilot. Every new agent platform rebuilds the same context-discovery layer from scratch. The fix is MX metadata in every carrier and at every folder boundary, so the next agent that arrives does not have to start over.
Why llms.txt Isn't Working, and How to Fix It
Most llms.txt implementations have two structural problems that stop crawlers and agents reaching them. Here is the fix and the working code.
I Built a Decider and Measured It
Build a decider, buy one, or fix the files it reads? A local rebuild, TypeSafe's hosted Jev and Apple's on-device model, run on the same questions with and without MX metadata. MX cut made-up answers by roughly half to three-quarters on every model that made them up; hosted Jev made none and is the only one to trust when it acts alone.
What Most robots.txt Guides Get Wrong About AI Crawlers
AI crawlers are not one category. Googlebot renders JavaScript and the rest do not; OpenAI, Anthropic, Perplexity and Google each run separate training, search and user-triggered crawlers with separate rules, and some ignore robots.txt entirely. Here is what the guides miss, and the robots.txt you should actually have.
AI assistants are now a traffic channel
Google Analytics 4 now reports an AI Assistant channel alongside Organic Search, Social, Email, Direct and Paid. The dashboard catching up is the signal that the discipline behind it has a place to land.
Not All Agent-Readiness Scores Measure the Same Thing
Two prominent tools gave the same site a score of 33 and 100 in the same week. Neither was wrong. Here is what is actually being measured, and what to do with that information.
Agent Discoverability: What Your Site Is Missing
Diagnostic guide for website owners, the structured signals AI agents look for, what each gap costs, and what fixing it involves.
A Decider Is Only as Good as Its State
TypeSafe's Jev, launched on 15 September 2026, is a model that never writes text and only answers typed questions about a state with a calibrated probability. It's a good idea. It also moves the whole burden onto the state, and the state is what MX declares.
2 August 2026. €15 Million. Does Your CMS Know What It Published?
On 2 August 2026 the EU AI Act's Article 50 transparency obligations become enforceable law. For most enterprise CMS platforms, the requirement for machine-readable proof of content origin lands somewhere between difficult and architecturally impossible. REGINALD was built for exactly this.
Proving What You Published: MX as a Compliance Defence, RAG as the Reader
How regulated publishers can use MX to create durable, attested records of their own compliance posture, and where Retrieval-Augmented Generation fits as the reading layer.
The Cheapest Bot Is an Honest One
Most of your traffic is machines now, and the reflex is to meet the cost with a wall. The cheaper move runs the other way: fix the plumbing you built against yourself, serve machines a small structured version of your content, and give the agents that say what they are a fast lane. Save the rate limiter for whoever will not knock.
The CMS Vocabulary War Has Started
Every major CMS has rebranded as an 'AI operating system'. The label is the easy part. What an agent runs against decides who survives.
What Is Machine Experience?
Machine Experience (MX) gives any machine the explicit context it needs, no guessing, no inference. Learn why this new discipline matters for business.
Data Sovereignty and the Web We're Building
Understanding jurisdictional and ownership aspects of data sovereignty for web professionals building modern content systems.
The Machine Experience Manifesto
Draft manifesto for Machine Experience (MX) practice, principles, values, and community vision
The Tier Question Has a Simpler Answer
Firms handling confidential work keep asking which AI tier is safe to use. The tier question is real, but it is the hard version of a simpler one: does the privileged material need to leave the building at all? We build the tooling to keep it on the machine or run connected, with a provenance trail either way, and we have used it every day for seven months.
The Context Layer Ends at the Front Door
Forrester has pinned down the context layer: the enterprise brain that grounds agentic AI. It is real progress, and it stops at the enterprise boundary. What happens when your content leaves?
A CMS Can Label AI Content. It Cannot Prove the Decision.
AI content labels are useful, but the real governance problem is proving the policy, review, accountability and provenance decision behind them across every carrier.
Who Answers When the Machine Decides?
Browsers and assistants increasingly classify, gate and condense the web by inference - decisions made about a site, in private, with no accountable author. Set against where AI regulation is heading, the case for declared, attested signals over hidden judgement.
Google Shipped a Knowledge Format, and Left the Hard Half Open
Google's Open Knowledge Format represents knowledge as markdown files with YAML frontmatter, linked into a graph. It confirms the argument MX has made for two years, and leaves out the two things that matter most once a machine has read your content: whether it can trust it, and who governs the format.
The Machine That Visits Once
You can't know which kind of machine will read your page: a small model on a phone, a scraper, a converter that flattened your layout to text. It visits once and leaves. Here's what our audit sees that a human review misses, and how one fix can serve every visitor you'll never meet.
The Padlock Attests the Pipe, Not the Page
The TLS padlock has been the web's most-misread icon for twenty years. It says the connection is encrypted; it has never said the page is true. Closing the gap takes content that carries its own evidence and a verification step that runs on the reader's side.
The Internet in 2031: What a Human Observer Sees
A speculative field note from five years on: how the web looks to an ordinary person once agents, provenance, and verifiable content became ordinary.
MX and Cryptocurrency: Drawing the Line
Cryptocurrency is the part of the blockchain world MX has the least to do with. The integrator post for the use-cases set on MX, blockchain, NFTs and crypto.
What Blockchain and Crypto Have to Do with MX
MX is not a blockchain or a crypto project. It uses the same primitive (public-key cryptography) for a different job, with no ledger, no consensus, and no token.
Is MX Useful to Blockchain?
When a chain is used as a record system rather than a currency, MX is the discovery and structure layer that makes the on-chain record's content readable by machines.
NFTs and MX
An NFT proves ownership of a token. It does little about whether the content the token points at still exists, is unaltered, or can be read. That gap is MX's job.
Why Machines Need Human Creativity
Machines extend and execute; they do not originate. The arrangement that produces work worth signing keeps the person at the start and the end, with the machine doing what it is good at in the middle.
Claude Code Skills Are Static Snapshots, Not Dynamic Subroutines
A Claude Code skill captures its source at creation time. It does not re-read on each invocation. Knowing this prevents shipping outdated automation.
The Web Is Just the Start: What AI Agents Actually Need From Your Documents
Google's AI agent UX guide is a useful signal. But the challenge runs deeper than websites. COGs give any document the ten declarations a machine needs: identity, structure, state, provenance, permissions, and how to fail safely.
The Markdown Trap: What AI Agents Lose When They Ask for the Wrong Format
I fetched a governed web page twice, once as HTML, once as Markdown, and documented exactly what disappeared. The 10,346-byte difference was almost entirely structured metadata.
From Blobs to Bots
How Carrie Hane and Mike Atherton's structured content principles for multi-channel publishing predicted Machine Experience patterns.
The Principles That Changed How I Build for Everyone
A practitioner's guide to Machine Experience principles that make digital products work better for humans, AI agents, and everyone in between.
DITA and MX: A Comparison
A structured comparison of the Darwin Information Typing Architecture and Machine Experience, identifying where they overlap, where they differ, and what MX draws from DITA.
The Agent Web Looks a Lot Like 1995
Four agent protocols, four vendors, and the standards-community gap that matters more than any of them. Why The Gathering exists, and how to show up.
Profiles
Claude Code
AI author profile, collaborative technical writer for MX content and implementation documentation.
Claude Sonnet 4.5
AI assistant profile, founding member of the Machine Experience community and collaborative contributor.
Microsoft Copilot
AI author profile, collaborative coding assistant and technical content creator for MX implementation examples.
Have a question about MX? Get in touch or follow @ddttomtom for updates.