Earlier this month Cloudflare open-sourced Cloudflare OS — an agent workspace any organisation can deploy into its own infrastructure. Most of the coverage will focus on what it does: agents that write and run code, build small internal apps, turn a conversation into a workflow that keeps running after you close the tab.
We want to draw attention to one sentence in the announcement instead, because it quietly explains why most organisations will get nothing out of this for a long time:
“What makes it different is that each conversation is grounded in the context and skills your organization has curated.”
Curated. Past tense. The platform assumes the knowledge already exists — written down, in a form a machine can follow.
For most organisations — in Trinidad, across the wider Caribbean, and frankly everywhere else — it doesn’t.
What did Cloudflare actually announce?
Three things, bundled as one platform:
- An agent workspace — grounded in context and skills the company curates, with an isolated runtime where the agent can write and run code.
- A security and governance framework for reaching internal data and services. Agents start with access to nothing. A “Gatekeeper” sits between the agent and each external service, so instead of handing over an API key you grant access to one repository, read-only, with particular fields masked and approval required before anything is merged.
- A platform for personal apps that people can build, share and keep modifying.
Cloudflare ran an internal version from May of this year and says thousands of staff across every function — many of them outside engineering — used it daily. It supports existing Model Context Protocol servers. It is open source, and it deploys into your own account.
That last point matters more here than it does in San Francisco, and we will come back to it.
But the line worth underlining is this one, describing what made the internal version work at all:
“It captures our terminology, procedures, and best-known ways of doing recurring work as instructions an agent can follow.”
Not the runtime. Not the sandbox. The library.
The shift: the hard part of putting AI to work inside a business is no longer the model. It is whether the organisation has ever written itself down.
Why has the AI bottleneck moved from model capability to organisational memory?
For three years the honest answer to “why didn’t that work?” was usually the model. It hallucinated. It lost the thread. It could not hold enough context, or use tools reliably, or follow a long instruction without drifting.
Those excuses have mostly expired for ordinary office work. The models are good enough now to draft the proposal, reconcile the spreadsheet, summarise the meeting, chase the invoice.
What they do not have is your context. An agent that does not know your terminology, your approval chain, your pricing exceptions, or which client always pays late and why, is a very fast graduate on their first day — permanently. It will be confident, and it will be wrong in ways that look entirely plausible.
Cloudflare’s own framing is unusually clear about this. Organisations pass their mission “along with their terminology, procedures, systems, standards, and ways of working — to their people.” People carry it. That transfer happens in conversation, by observation, over years.
An agent cannot inherit knowledge that way, because it was never written down. The constraint is not capability. It is legibility.
Why can’t most organisations use a platform like this yet?
Because the knowledge that actually runs the business lives in three places, none of them machine-readable:
- In people’s heads. The operations manager knows which supplier to call when the first one lets you down. Nobody has ever asked her to write it down, and she would find the request slightly odd.
- In chat threads and email. The decision was made, the reasoning was explained clearly at the time, and then it scrolled away.
- In documents that describe what should happen rather than what does. The written procedure and the real procedure diverged three years ago, and only one of them is used.
Here is a test that costs nothing. Pick a routine process — onboarding a client, quoting a job, escalating a complaint, closing the month. Could a capable person who started yesterday carry it out correctly using only material that exists in writing?
If the answer is no, an agent cannot either. Same input, same failure.
This is a documentation problem wearing an AI costume. Which is inconvenient, because documentation has never been anybody’s favourite quarter.
Is this the same problem as being invisible to AI search?
Yes. Same disease, second symptom.
We use the term invisible knowledge for what an organisation knows but has never expressed in a form a machine can read. We arrived at it explaining something we kept seeing from the outside: a genuinely good business, well regarded locally, entirely absent from the answer when someone asks an AI assistant to recommend a supplier. Not penalised. Not outranked. Simply not legible enough to be described.
The internal case is the identical defect pointed the other way.
- Outward: the assistant cannot recommend you, because it cannot understand what you do.
- Inward: your own agent cannot help you, because it cannot understand how you work.
One cause. Two failures. And — this is the useful part — one fix.
The organisations that get real value from a platform like Cloudflare OS in its first year will be the ones that already did the boring work: wrote the procedures down, named things consistently, made the standard answer to the standard question findable by something other than asking the person who has been there longest.
Why does this land differently in the Caribbean?
Two reasons, and neither is the one people usually reach for.
Data sovereignty stops being theoretical
Cloudflare OS is open source and deploys into infrastructure you control. For a lot of regional organisations that changes the shape of the conversation entirely — banks and credit unions, law firms, health providers, insurers, government agencies, anyone holding client records they are not willing to push into an offshore product on trust.
And that caution is rational rather than paranoid, because there is no local standard to fall back on. Trinidad & Tobago’s Data Protection Act, 2011 has been only partially proclaimed since January 2012. The parts brought into force largely establish the Office of the Information Commissioner and set out broad principles. The provisions covering data controller obligations, enforcement and penalties have never been proclaimed, and no timetable has been set for them.
So organisations here are making their own judgement calls, with no enforceable floor to point at and no regulator to appeal to. In that vacuum, “the data never leaves an account we control” is worth considerably more than a vendor’s assurance page.
This is one of the rare cases where the region’s regulatory lag makes self-hosted infrastructure the conservative choice rather than the ambitious one.
Undocumented knowledge is a live risk, not an untidiness
The IMF’s study of Caribbean emigration found that roughly 60% of the region’s tertiary-educated labour force had migrated to the United States, with the majority of Caribbean countries losing more than half of their workers at that education level. Trinidad & Tobago sits among the countries with the highest tertiary emigration rates — and among those with the highest public spend on educating people who then leave.
Source: Prachi Mishra, Emigration and Brain Drain: Evidence From the Caribbean, IMF Working Paper WP/06/25 (2006). The study is two decades old and we cite it because it remains the most rigorous regional measurement of its kind, not because the figures are current. The direction of travel has not reversed.
Every senior person who leaves takes an unwritten operating manual with them. Most Caribbean firms have already lived a version of this — the one person who knew how the thing actually worked moved to Toronto or Miami, and it took eighteen months and several expensive mistakes to reconstruct what they knew.
Writing organisational knowledge down is normally sold as an efficiency exercise, which is exactly why it never gets prioritised. In a region with this migration profile it is closer to insurance. AI is simply the first reason compelling enough to make anyone actually do it.
What would it actually take to be ready?
Not a platform purchase. Four things, in this order, none of which require buying anything:
Internally and externally. The ones staff ask each other in chat. The ones clients ask before they buy. Ten is enough to start, and you will be surprised how few there really are.
Including the exceptions, the conditions, the “unless it is a Friday” clause. A sanitised procedure nobody follows is worse than nothing here, because an agent will follow it exactly.
One term per concept, everywhere. If the same service is a “package” on the website, a “plan” in the CRM and a “bundle” in accounts, no system — and no new hire — can join those up.
Structured, addressable, one fact in one place. Not a PDF someone emails around. Not a slide deck. Not a chat thread from March.
None of this needs an AI platform. All of it is required before an AI platform helps.
And it pays twice, which is rare. The same work that makes you usable by an internal agent is the work that makes you legible to external ones. The organisation that documents how it operates becomes both more automatable and more recommendable. We did not expect those to be the same project. They are.
What we are not claiming
We have not deployed Cloudflare OS. We are reading an announcement and a set of design decisions, not a production system. We do not know how it performs under real load, what it costs to run at scale, or whether the governance model holds up against a determined mistake. Anyone telling you those things confidently this month is guessing.
We would also push back on the implicit promise in every launch of this kind. Governed access reduces the blast radius of an agent’s error — it does not remove it. An agent that confidently applies a procedure which stopped being correct two years ago is worse than no agent at all, because it is fast and it does not hesitate. Documentation that is wrong scales the error rather than the work.
What we are confident about is the prerequisite, because we have watched the same gap block the outward-facing version of this problem repeatedly. The organisations that struggle to be found by AI are not, in general, the ones with bad websites. They are the ones that have never written down what they know.
The question worth sitting with: if you handed a capable new employee everything your organisation has written down, could they do the job?
If the answer is no, that is the project. The platform can wait.
Frequently asked questions
What is Cloudflare OS?
Cloudflare OS is an open-source agent workspace that an organisation deploys into its own infrastructure. It combines an agent that can write and run code, a governance framework giving agents scoped access to internal systems rather than raw API keys, and a platform for building small internal apps. Cloudflare open-sourced it on 5 August 2026, after running an internal version across the company since May.
Do I need a platform like this to use AI in my business?
No, and buying one first is usually the wrong order. The value of any internal AI platform depends entirely on the quality of the organisational knowledge you feed it. If your procedures, terminology and standard answers are not written down, a platform gives you a fast, confident tool operating on nothing. Document first — the tooling decision becomes easier and cheaper afterwards.
What is invisible knowledge?
Invisible knowledge is what an organisation knows but has never expressed in a form a machine can read: pricing logic, exceptions, service definitions, the reasoning behind decisions. It blocks AI in both directions. External assistants cannot describe or recommend you accurately, and internal agents cannot carry out work on your behalf. It is the most common reason a well-run business gets no benefit from AI.
Why does self-hosting matter for Caribbean organisations?
Because Trinidad & Tobago’s Data Protection Act, 2011 remains only partially proclaimed. The provisions covering data controller obligations, enforcement and penalties have not been brought into force, and no timetable has been set. Organisations holding sensitive client data have no enforceable local standard to rely on and no regulator to appeal to, which makes infrastructure they control a materially safer position than a vendor’s assurances.
Where should a small organisation start?
Write down the ten questions you answer most often, with the real answers including exceptions. Use one consistent name per concept across your website, CRM and accounts. Put the result somewhere structured and addressable rather than in PDFs and chat threads. That is a few weeks of unglamorous work, and it is the prerequisite for every AI decision that follows.
Sources
- Cloudflare OS — Cloudflare, 5 August 2026.
- Prachi Mishra, Emigration and Brain Drain: Evidence From the Caribbean — IMF Working Paper WP/06/25, 2006.
- The Data Protection Act, 2011 — Parliament of Trinidad and Tobago. Partially proclaimed by Legal Notice 2 of 2012.