📰 newsreader

There was a time Cloudflare just made the internet better by staying hidden like Batman’s identity: protect & fight the bad people, for the sake of the global city of the Gotham… err I mean the internet.

When I installed Cloudflare on my website for the first time 10 years ago - it saved me tons of megabytes, saved me money on bills and also the moment it sent me a monthly reporting on how my site was performing… that’s when I knew Cloudflare (CF) was AWESOME.

Because Cloudflare was doing a few things well: it sat in front of your site, ate the attacks, cached the static stuff, did the DNS, and that’s it.

No BS.

Good infrastructure. Fast. Reliable. Boring in the best way.

The new stuff. Not so good.

More importantly this is my personal opinion on my personal blog, you can disagree, that’s ok with me!

Fundamentally I work at a small AI startup, rely on CF and I am a not-totally-unhappy-but-also-not-totally-happy

customer of theirs.

CF is a big business, a big money making machine. Today they are bigger than ever and still route a ton of the daily web (something like 1 in 3 requests or so).

Great job for the shareholders and whatnot, the safety and caching layer they sell is doing something right and pays the bills (stock is at an all time high as I write this).

No matter how well the stock is doing the last few years CF turned itself into something a little more cringey and clique (but not as bad as others cough △).

Today it doesn’t feel like a company ran by good engineers, it’s now PMs and vibecoders with what I call AI psychosis: dream it -> vibe it -> ship it.

Lost touch with reality

First of all tons of more outages than ever, remember that React useEffect fkup [0]? Complete insane that this would happen at an infra company that runs a third or so of the web.

Outages are omnipresent but let’s talk about my biggest complaint: the DX.

The developer experience feels like bolted on as an afterthought.

But CF is an infra company, with badass engineers, shouldn’t the DX be top?

Instead CF decided it also wanted to be a cloud platform for everyone, kids, dogs, and vibers.

How did we get there?

The AI product manager mindset.

Ship slop, post it on X, get 200 like rinse and repeat.

Instead of focusing on expertise, reliability and simplicity…

This is how an INFRA company got product-managed into something closer to the problem it used to solve.

When an infrastructure company gets crushed by a full product management org, you know what’s coming next: first features multiply. Naming turns into marketing (Hyperdrive? sounds cool right? ship it). All that seem to matter are some half-finished primitives, cousin-wife style.

Examples? Ok -> Let us look at data storage.

They got D1 (SQLite serverless), Durable Objects with their own SQLite, KV, R2, Queues, and Hyperdrive to speed up external Postgres or MySQL.

STILL no real first-class managed PostgreSQL that feels native. Keyword here is that it must feel native.

Hyperdrive is a smart connection pooler and cache for databases that live somewhere else.

Useful but an admission they never built the database most serious apps still want.

In the end you end up gluing three or four storage products together and hoping the docs for that combo aren’t six months out of date (IYKYK).

Ok maybe you’re gonna think: “YOU DON’T KNOW WHAT YOU’RE TALKING ABOUT DUDE”.

Let’s take a look at the compute side.

Woops.

Compute is the same mess.

There’re Workers.

There are also Dynamic Workers, runtime-spawned isolates sold as a light alternative to containers for AI agents and untrusted code.

Then Sandboxes (on Containers).

Then full Containers…

and a bunch of “code mode” paths around it all, so agents can write and run stuff.

Each one has different isolation, startup time, pricing, and bindings.

None of them is just “the place you run code.”

Picking between them means reading multiple docs pages that contradict each other or lag the actual product.

Ok maybe you’re gonna say: “ DUDE WE NEED DIFFERENT LAYERS OF COMPUTE FOR DIFFERENT”. OK THEN - I MUST BE WRONG AGAIN.

Let’s look at the latest and greatest hype.

AI Agents.

Agents are even noisier.

Agents SDK.

Flue.

Project Think.

Cloudflare OS (they just open-sourced their internal agent workspace).

And the observability gets bolted on later (more on this below this section).

Every new announcement adds another harness or framework instead of finishing the one they already have.

Classic product manager move: add more surface area, maximize incoherence till the point developer experience feels like an internal experiment that escaped the sandbox (get it?).

RAG same story. AutoRAG got renamed AI Search. It’s just a managed pipeline on R2, Vectorize, Workers AI. Fine for demos and hackathons.

But in real use it lags proper RAG platforms or even a decent open-source stack on quality, filtering, hybrid search, and actual visibility.

When it “Works for simple cases”… you know that is a low bar for an infrastructure company.

Docs kill me. You can’t just ‘redesign’ the UI and call it a job done. Pages ship incomplete. Examples rot. New products show up without the precise, versioned reference material (or even basic SKILL.md stuff for agents) that real infrastructure needs.

Trust goes away When the company itself is shipping AI-generated stuff that later needs disclaimers and TODO cleanups. [1]

Every engineering knows that documentation is part of the product. If you can’t get the docs right - how can you get adoption and trust? Treating it like secondary marketing…no bueno.

Peeking at Workers AI, the inference layer, just keeps the pattern going:

As latency got better and they added bigger open models, things seemed going up, then you look deeper and you see it still trails specialized providers on speed for a lot of workloads and on having the newest frontier models.

Cloudflare pitch itself as the place to run agents, but the catalog is worse than last year’s IKEA.

Then performance force a lot of teams to send the hard inference somewhere else and treat Cloudflare as plumbing again. [2]

A lot of this comes from caring more about fighting Vercel on X than building a solid layer on top of the infrastructure layer.

Edge functions, frameworks, agent runtimes, “full-stack” announcements keep coming day in and day out, but the core network, reliability, and simplicity that made Cloudflare different get less consistent attention. It feels like AI psychosis has wrapped the whole upper echelon, so the product org gets measured on feature velocity and competitive posts instead of making the existing stuff excellent.

This AI craze led the whole team to produce a sprawling catalog of overlapping tools, each good enough for a blog post. None of them the clear durable answer an infrastructure customer actually needs.

This is what product managers do to an infrastructure company when nobody stops them.

They optimize for announcement cadence and surface-area growth.

The coherent, trustworthy infrastructure layer that made Cloudflare matter? That layer is getting harder to find under all the noise.

Cloudflare has spent years announcing progress on Observability. Logs became “Generally Available.” Unified Observability section showed up in the dashboard. Automatic tracing went into open beta. Query Builder, metrics views, OpenTelemetry export got added. Marketing says first-party observability that finally matches the platform.

SIIIIIIKE

Reality feels a lot different because the core pieces stay partial, buggy, or just missing when you need them.

Their own docs list hard limitations that never went away. Tracing is still open beta [3] (it literally says BETA on the badge next to the name Tracing on their docs site AUG 2026).

Non-I/O operations often report 0 ms because of Spectre mitigations in the runtime. Trace context doesn’t propagate to external services, so end-to-end visibility across non-Cloudflare stuff is broken on purpose. Span attributes are incomplete; they’re still planning to add more. Even paid accounts reported the platform wrongly applying aggressive 1% sampling even when head_sampling_rate was set to 1 and usage was way under quota - luckily it got fixed [4]. How can you ship logs with bugs, does anyone else use observability?

Community threads show the day-to-day pain (don’t worry they fired the community managers facepalm one more bad decision): people report that logs vanish from the dashboard while wrangler tail

still shows them in real time. One big March 2026 incident had most logs stop showing across an entire account—only scattered entries appeared even with correct config, available quota, and no code changes.

One developer put it clean: “logs and traces from workers in one click is the missing piece, workers observability has always been the weak spot.” They keep shipping adjacent features while the basic experience of “I want reliable logging for Workers” stays incomplete.

Remember this scenario? it’s Monday. Open HN and shout “Dad wake up new Cloudflare product dropped”. Within hours a Product Manager posts the announcement on X. Post is optimized for engagement—clean screenshots, bold claim, a thread. Then the long blog post: a 6000 word story, aspirational framing, dashboard screenshots, a few customer quotes, and almost no real technical talk about the hard trade-offs or why they picked this design over the alternatives. Is talking about engineering decisions considered secret sauce? Or was it because ChatGPT suggested it?

Every other Monday a new drop:

Another agent harness.

Another sandbox variant.

Another managed RAG pipeline.

Another way to run code.

The post smell like marketin from a mile away, but you’re an infrastructure company. Who else is gonna read those blogs? Dieticians?

The deeper problem sits above the product managers. Look at the current leadership and the question is unavoidable: where are the infrastructure people who once defined Cloudflare?

They still have strong engineers (hi Kenton Varda). But the center of gravity shifted. CF is optimized for announcement cadence, competitive positioning against Vercel, and the next AI story.

The people who understand the hard constraints of running a global anycast network, the subtleties of isolate isolation, control-plane reliability… seem to have less authority now.

You know how you know? The 2026 workforce cut made the priority clear. About 1,100 roles gone framed as an AI-driven (AI Slop) move toward an “agentic” operating model. Old talent keeps flowing to OpenAI, Anthropic, and other AI labs.

What Cloudflare needs is more T-shaped infrastructure peepz. People deep in systems, networking, runtime design, and reliability. And product managers should exist to serve those people.

You might not believe it reading this far but a lot of people like me still love the underlying product. The cloud network. That’s exactly why the current path is so frustrating. I feel like this is death by a thousand AI slops: AI Psychosis.

Cloudflare still has time to reverse it. Hire people that care and don’t just chase the F.I.R.E. at 35 with multiple hundred thousand dollars a year + RSUs for posting a viral tweet.

[0] https://blog.cloudflare.com/deep-dive-into-cloudflares-sept-12-dashboard-and-api-outage/

[1] https://www.thestack.technology/cloudflare-matrix-blog-ai-assisted-vibe-coding/

[2] https://developers.cloudflare.com/ai/models/

[3] https://developers.cloudflare.com/workers/observability/traces/

[4] https://www.answeroverflow.com/m/1484012295314215104