I love the problem definition and sketch of the solution space here. In some sense the "dumb network" footprint is what both TEFCA and CMS Health Tech Ecosystem are moving towards (networks facilitating trust and routing, but smart edge nodes that can talk directly to each other). I'm curious if you think this is likely to succeed in (re)shaping networks, or if not, why not?

Made using freely available FHIR APIs from open.epic.com.

Yes, the best frontier agents right now output undecipherable language and it's annoying -- but a compensatory corollary: they will happily accept your own made-up language as input, where you obliquely refer to design ideas, architectural components and goals with the laziest linguistics gestures.

Some contexts are just cursed? Reading through the thick Claudish: "The last one is the most useful failure: the model can be set and the card still refuses, so the gate isn't the value — it's a dirty flag only the real picker sets." Bottom line; model returns to say "I failed" instead of solving the problem (in the same way we've already established will work, like, 10 times by now).

So, out of the box Android migration is... decent. But you lose settings for lots of apps, and you lose all non-Play-store apps. But with wireless ADB and Claude code coordinating, I got *everything* reconfigured. Including my growing pile of vibecoded apps + my open source podcast db.

Can y'all please settle on a direction of travel for usage progress bars @OpenAI and @AnthropicAI?

"/goal [briefly described exit condition]" is basically asking the genie to grant your wish... with all the, uh, provisos, and quid pro quos Disney warned about.

As you know, ONC can require any triggers it wants whenever/ifever they mandate support for the patient data feed. Historically when an IG defined multiple optional capabilities, ONC has designated their own set of mandatory ones to align with their priorities. Reconciliation with the published IG can happen (much) later or never. So I don't see lack of Argo-mandated triggers as a fundamental flaw. (Remember this spec came out of HTI-2 proposals that were themselves quite vague on use cases -- so instead of trying to pick winners or downselect data types, we described a general purpose approach that covers any uscdi patient oriented data. This should allow what you describe as "enrollment" to the extent that it is supported by uscdi -- and if it is not supported by uscdi then there is more fundamental work needed before onc can try to solve the problem.)

"The federal contracting demimonde may feel odd to the uninitiated civilian eye (including me), but is ultimately a fixture of the GovCon firmament."

Some harnesses seem to actively penalize thorough reading (i.e. agent tries to read a file; harness is like "here's 50 lines, if you reeeally want more, you can make another read call".)

Totally agree it's about "how easily you can let it go." This is why I don't use agent frameworks and rewrite (re-prompt) infra I think I might need over and over again for ~every project. (Seriously: keeps infra light and keeps me grounded in what I actually need vs can jettison).

You don’t need to be a cardiologist to see it – it’s obvious… ex post But we’ve been looking at ECGs for predictors of sudden death for 100 years No one saw it until the AI pointed it out!

opus 4.8 with the fable context is some real flowers for algernon shit

I am *constantly* typing some variation of "the output should be self contained and not make references to our conversation here -- the reader will not have this conversation in context"

With W3C Digital Credentials API, it's *finally possible* to design a health-data-sharing protocol that's idiomatic (standard FHIR-based resource sharing, questionnaire filling, etc) and works cross-platform (from web and mobile app requests to Android + iOS wallets)! See article for background + details on my updated "SMART Health Check-in Protocol".