Dagens Vibes — 24. juli 2026

Armin Ronacher ⇌@mitsuhiko

We've recently made Pi's cache behavior more visible. This site has been debating whether agent harnesses are helping or quietly torching their caches. That seemed like a good excuse to explain how KV caches actually work and how Pi helps (or doesn't).

Mark Ajzenstadt@mardehaym

A PE-backed healthcare billing platform had two people hand-writing every rule in their claims system. 300+ denial codes, payer variations that change three times a year. We built seven production AI agents that handle it. Four months. Zero PHI exposure. I want to walk through what that actually took, because "we added AI to the product" tells you nothing. Their system processed real claims under real HIPAA constraints. A wrong output costs money or violates a regulation. This is production healthcare, not a demo. First month, we didn't pick models. We mapped data. Where is it, what are we NOT capturing, how do we structure it so agents reason correctly? Most teams skip this. They bolt a model onto the product, watch it hallucinate over unstructured inputs, and conclude AI isn't ready. We built an enrichment layer. 34 dynamic variables assembled per claim before any LLM touches it. Pre-computed, versioned, auditable. The agent doesn't search for context. It receives clean, ranked facts. Then we designed the architecture. Every agent follows a fixed pattern: pre-compute context, strip all patient data before the model sees it, call the LLM, validate output against a strict schema, let deterministic code decide the action. Our engineers wrote validation logic that rejects any output outside a bounded allowlist. If the model suggests something the allowlist doesn't cover, the system fails closed. Seven agents, each locked to a single workflow. One handles denied claim follow-up. Another runs billing reconciliation. Each has its own enrichment payload and output schema, validated by a dedicated eval suite. Their PE operating partner drove AI transformation across the full portfolio. The other portfolio companies now benchmark against this agent system. You bolt a chat widget onto your product and call it intelligent. Production agents need enrichment pipelines pre-computing context and eval harnesses catching regression. They need audit trails and human-in-the-loop controls at confidence thresholds. We spent more engineering hours building that scaffolding than integrating the model. We route calls across two model providers. Swapping one for the other changes nothing in the output. Model integration was the shortest line item in the build.

Domenic Denicola@domenic

I wrote about my current agentic coding setup. If you're still leaving your laptop lid open, using raw Claude Code, this is for you!

OpenAI@OpenAI

ChatGPT Voice is now in the desktop app. Control your computer and direct multiple agents running in ChatGPT Work or Codex, using just your voice. It's powered by GPT-Live, so it can speak, listen, and coordinate work in the app at the same time. Rolling out globally today on macOS and Windows to Plus, Pro, Business, Edu, and Enterprise plans.

Simon Last@simonlast

Uncle bob is agent-pilled!

Uncle Bob Martin@unclebobmartin

I’m significantly older than you. I started coding in the late 60s. My current strategy is to not read any of the code written by my agents. That’s the only way I can take advantage of their productivity. What I do instead is to surround the agents with extreme constraints. Unit tests, gherkin tests, QA procedures, quality metrics, mutation testing, test coverage, and a plethora of others. In the end, I have very high confidence in the code they produce because they’ve had to run the gauntlet of all of my constraints and tests.

♥ 8.633↻ 770💬 273🔖 5.370

Chris Tate@ctatedev

Now available: accessibility audits agent-browser a11y <url> → WCAG violations → Selectors + fix links → Powered by axe-core