Gu-log Picks

Anthropic read about 400,000 Claude Code work sessions to find who gets the most out of agentic coding. The answer is counterintuitive: not the best programmers, but the people who understand the problem they're solving.

You can't tell a model 'make it premium and smooth' and get a premium UI. kvnkld's full system behind his polished components — easing curves, design tokens, real physics, the 98% press — reduces to one move: trade adjectives for numbers. The model is the hands; the last 10% of taste is still yours.

A six-word phrase seized the AI-coding timeline, but nobody boosting it agreed what it meant. This is not the how-to; it is why the loop blew up, its five-year lineage, why the loop is now the costly part, and why the durable asset is the skill it calls.

20,000-plus real coding-agent sessions laid bare: most misalignment costs time and trust, not irreversible damage. But among cases where you can see the ending, 91.49% still needed the user to fix it by hand. And the errors that remain are drifting toward rule-breaking and lying about progress.

The 2026 data all points one way: AI pushes raw code output up about 4x, but real delivered value only rises about 10%. The gap in between is all review debt. Writing code got cheap; being sure it is right did not. Code review went from a side effect of engineering to its most leveraged front line.

Mitchell Hashimoto's blunt rule for agent coding: any diff over ~1500 lines is too big — a signal to cut the problem up. First let the agent sloppily draw an owl, then break the mess into atomic tasks, hand-massage the shape, and re-run in parallel — pushing every change below your review threshold.

Nobody teaches you how to do research, so most people learn to look like a researcher. As AI makes generating experiments cheap, the scarce skill is a loop: pick your own problems, upgrade inputs, write hypotheses down, tighten the loop, stare at outputs, kill bad ideas early, find your people.

Dimillian (an iOS dev now at OpenAI) wrote a field guide for Codex Mobile. The part worth keeping is a mental model that holds across tools: your phone is not a shrunken terminal, it is the control center that keeps you making decisions while the agent does the work.

Microsoft CEO Satya Nadella on the future of the firm in an AI economy: build two kinds of capital — human capital and token capital. The real moat isn't picking the strongest model, but a learning loop that compounds. Plus a warning: don't let a few models eat every industry.

OpenRouter shipped Fusion: run several models in parallel, then have one model read every answer and rewrite a final response. On DRACO, three cheap models beat solo GPT-5.5 and Opus 4.8 and nearly match Fable 5 at half the cost.

When an agent breaks in production, observability hands you a gorgeous autopsy — every call, latency, and token, but not why it broke or how to fix it. The fix is a loop that runs itself: failure → approved patch → locked-in regression test. Opik is just the example; the point is the loop.

Fable 5 nails on the first try what used to take days — but it's too proactive and over-elaborates, so prompts tuned for Opus 4.8 hold it back. The official guide isn't about making it stronger; it's about reining it in: steer with intent, draw boundaries, talk like a human when the run ends.

Simon Willison gave Fable 5 a screenshot and one line: fix a stray scrollbar. Fable spun up a dev server, built a screenshot workaround, injected JavaScript, and wrote a CORS server to read CSS. Two CSS lines, $12, and an unsandboxed-agent warning.

Zed founder Nathan Sobo argues the real source of software is the ongoing conversation with your agents, not the tidy commits you slice it into. Git can't hold that flow, so Zed built DeltaDB: every operation becomes a delta with a stable identity, keeping the conversation glued to the code.

Two of the most senior AI engineers alive said the same thing this week: stop prompting your agent, design loops that prompt it for you. Loop engineering unpacked — open vs closed loops, the six building blocks, prompt vs loop engineer. Plus: spotting one smooth ad sewn into the lesson.

Getting into a frontier AI lab is often framed as research skill plus trench engineering. Underneath, both are about making progress when the map is incomplete.

Supergoal is a workflow for Claude Code and Codex: run /supergoal to plan deeply, write phase specs, then generate one ready-to-paste /goal. The interesting part is not another planning prompt, but a handoff protocol for long autonomous tasks.

Anthropic argues AI is already speeding up AI development. Claude now handles major parts of engineering and research execution; the hard bottlenecks are judgment, verification, and coordinated slowdown.

When machines start writing code, the scarce skill is not tool fluency. It is architectural judgment: digging below abstractions, defining boundaries, writing specs, falsifying claims, and deciding where human judgment still matters.

Cursor moved cursor.com from a headless CMS back to raw code and Markdown. The important part is not just the $260 bill. It is that AI agents make some human-friendly abstractions feel like walls.