My inbox has been filling with privacy-policy updates for two months. Most people read those as noise. I read them as a market signal. I started out in agency work, on the seam where a client's words meet the legal consent language before a feature ships. Reading that fine print out of curiosity, then going to find out what it means, has been a habit since long before this, especially when it affects other people. That habit is most of how I got here.
The signal is consolidation. Anthropic is pulling its products into one cloud surface. Memory got shared across chat and Cowork on August 25, the window to export your old memory closed on September 9, and on September 16 chat and Cowork started merging into a single conversation. I mapped the surfaces themselves separately; the point here is the shape of the move. It's the pattern the big consumer platforms ran years ago. The data that improves a product tends to be your data by default, and you opt out only if you go looking. The default is the strategy, and almost nobody changes a default.
The consumer product runs one loop for everyone. The developer product is where the value actually translates, and it needs something the consumer version doesn't have. It needs competency with your own business data and infrastructure, so the model's pattern recognition works in a closed loop for you.
Take Anthropic's own telemetry. Claude Code can export cost and usage data through OpenTelemetry, and that's documented. What the enterprise tier sells is the layer on top. The aggregation, the dashboard, the org-wide view of what your spend and activity actually are.
I built that layer myself on the Max plan. I took the developer API and the Claude Code harness with its hooks, pulled the telemetry into Claude Code, then carried it across Cowork and chat so the same numbers show up wherever I'm working. That meant rebuilding a single software feature as a system. Skills and a plugin and system prompts across each surface, the memory system holding state, and the orchestration glue tying it together. Once the numbers were in one place, my own cost turned out more interesting than I expected, and reading my actual spend up close is what let me correlate the micro to the macro. What it costs me to run this is a small, legible version of the cost curve every business is about to meet at scale.
None of it is locked to one vendor. There's an open-source layer underneath and local models you can run yourself, and the same setup gets you the same output. It takes heavy customization, and the work is knowing which parts are worth building yourself and which to leave alone.
The dashboard isn't the point. Observability and a verification loop are how you control your own destiny as a business instead of taking the vendor's word for it. That's the first thing I'd want in place before scaling AI anywhere.
That's the work an AI engineer should be doing. Creative problem-solving, architecting with the macro fundamentals in view, staying model-agnostic so you're not tied to one vendor's roadmap. When a recruiter or a hiring manager or the person setting an org's AI strategy isn't evaluating for that, the setup is already behind. It gets more expensive at the next cost shock, when the business has to scale on whatever it built.
I got here the way I get anywhere. I was never fine with the first answer a model handed me. There was always more context to push it with and a better outcome on the other side. If I'd stopped at the first answer, I wouldn't have built any of this.
Frequently asked
- What does owning your observability actually mean in practice?
- You pull the telemetry a vendor already emits into a place you control, wire a verification loop on top, and read it yourself. For me that was cost and usage from the Claude Code harness, carried across chat and Cowork so the same numbers show up on every surface I work in.
- Isn't this just the enterprise telemetry dashboard?
- The dashboard is what the enterprise tier sells. The telemetry underneath is already emitted and documented, so what you pay for at the top tier is the aggregation and the org-wide view. I wanted to understand that layer well enough to own it instead of rent it.
- Doesn't building it yourself lock you into one vendor?
- Not if you build it that way on purpose. There's an open-source layer underneath and local models you can run yourself, and the same architecture gets you the same output. The vendor is substrate, not a dependency.
- What consolidation are you pointing at?
- Memory got shared across chat and Cowork on August 25, and the two started merging into one conversation on September 16. I mapped the surfaces in a separate piece; here the pattern matters more than the specifics.
- Is it worth building yourself, or just paying for the dashboard?
- It depends on whether you need to trust the numbers. Paying is faster; building it means you know exactly what's measured and can wire your own verification on top. For anything you're scaling, I'd want the second.