Omasiri

Multi-agent coding gets safer when each role has less authority. The implementer cannot approve. The reviewer cannot edit. QA can file defects but cannot fix code. A merge needs independent review, compliance evidence, CI, and a readback. The human still unblocks agents and checks the work.

A timed-out EHR write may have succeeded after the response disappeared. The recovery path needs to prevent a second referral while it reconciles the first. www.omasiri.com/writing/how-...

Public-source experiment: run Gemini's new saved “skill” on one task across three changing inputs. Compare edits, omissions, and recovery, not just the first polished result. blog.google/products-and...

Opus 5.5 refused to enter credentials in a UI, even in my test/QA environment. That leaves an agent capable of testing the app stuck at the login screen. Protecting production secrets makes sense. Treating test credentials the same way makes legitimate automation less useful. @anthropic.com

A 53-second setup video for Jev tells me more than an architecture diagram: what the first successful path looks like, which command to copy, and where the setup breaks. That’s enough to decide whether the tool deserves a longer look.

There are several issues with the arguments in this paper; and I will post the gaps and counter arguments. arxiv.org/html/2606.05...

Model providers retire versions on their own schedule. What helps when one your feature depends on gets retired: a frozen eval set you can run in one command, an owner for the pin, and a swap you can roll back. https://www.omasiri.com/writing/how-should-you-handle-an-llm-model-deprecation/

The Verge reports drivers unknowingly enabled GM's Smart Driver tracking through OnStar, a connected-car service. Try labeling that button "Share my driving habits with insurance data brokers." Harder to mistake for a car feature. www.theverge.com/column/99417...

Acer says its new Vero 16 laptop opens without tools and has a replaceable battery. Nice. Can someone buying it in 2027 still order that battery in 2031, with a price and an Add to cart button? That would make the repairability pitch easier to trust.

Most teams argue RAG vs fine-tuning before they have the eval set that would settle it. The real split is where the knowledge lives and how fast it changes. A practical way to decide: https://www.omasiri.com/writing/how-to-choose-between-rag-and-fine-tuning/

Most agent budgets start with a model price and stop there. The bill is the loop: retries, tool output, context that grows on every step. I budget per completed task. Here's how to estimate before the invoice: https://www.omasiri.com/writing/what-does-it-cost-to-run-an-ai-agent-in-production/

Patient matching is a useful test of health software maturity. Duplicates, review queues, and merge rollback decide whether the product can handle real clinical data without pretending every record is clean. https://www.omasiri.com/writing/how-should-healthcare-software-handle-patient-matching/

Microsoft's $329B in future data-center leases is a weirdly useful AI metric. It says a lot of the roadmap has already become buildings and power bills before most customers know which features they will keep.

Microsoft 365 Copilot pricing getting regulator attention is exactly why AI bundling needs boring receipts. If renewal jumps, the screen should say what changed, what AI feature was added, and where the cheaper non-AI plan lives.

Claude, Anthropic's AI chat app, having shared chats show up in search is one of those product words doing too much work. If a button says Share, users hear "send this to a person." If Google can find it, the product shipped a publish button with softer lighting.

Apple reportedly slowing AI smart glasses over camera privacy is the least surprising product call of the week. A wearable camera turns every room into a permissions UI, and nobody wants to debug trust after the indicator light becomes a meme.

Watching 33 Raptor engines light on a Starship booster makes a lot of software launch debates feel very indoors. Hardware teams do not get to argue with the blast radius. The system either survives the ugly minute or it doesn't.

I wrote this because "we integrate with the EHR" hides a lot of product debt. For healthcare startups, the useful question is when the workflow has earned that integration, especially writeback. www.omasiri.com/writing/when...

ChatGPT now has Work mode, Codex mode, Ultra mode, Sol/Terra/Luna, and effort levels. I know why every knob exists. I also know most teams will end up making office folklore about which one to use for the monthly report.

I get the PromptQL pitch. Most companies already have a shared brain; it is just scattered across Slack, docs, dashboards, and the person everyone interrupts.

Local-first AI tools are getting more interesting to me than cloud meeting bots. If a tool hears roadmap arguments and customer context, send it all to us is a pretty big product ask.

The Fable 5 game demos are wild. I keep looking at the boring part after the clip: tuning controls, crashy edge cases, and sitting with a player while they say the jump feels bad 40 times. Still a very human way to spend an afternoon.

Piece on the dev loop I keep coming back to for coding agents: shared worklist, claimed stories, PRs, separate review, compliance, and reviewed status changes. Autonomy feels less weird when the workflow leaves evidence. www.omasiri.com/writing/maki...

The problem with people like Donald Trump, JD Vance, and Elon Musk, is they lack the self-awareness and familiarity with decency to be ashamed after shockingly inappropriate or rude behavior.

Microsoft's Majorana 1 chip uses topoconductors to create a new state of matter. They say it's more stable and reliable so, the path to a million qubits is now clearer.

Real estate tokenization is heating up—imagine owning a piece of Real Estate via blockchain! It sounds like the future of Real Estate investments.