An Exploded View publication

Reading Room

Vol. 1 · No. 1 Saturday, July 4, 2026 aikansh.com

This week's theme

Agents get a craft as the hype cools

AI agents earned real engineering discipline this week — a name, skill files, ranked tools — even as companies quietly rehired the workers automation was supposed to replace.

A 4-minute read · 8 stories

In this issue

01 Front Page

'Loop engineering' goes mainstream as agent iteration gets a name

Posts by the creators of Claude Code and OpenClaw gave a discipline a name — and sparked a field-wide conversation about how AI agents should actually be built.

Spotted on X: "loop engineering" — the practice of designing AI agents to iterate reliably on software problems — became this week's most-discussed phrase in AI development after posts by Boris Cherny, creator of Claude Code, and Peter Steinberger, creator of OpenClaw, went viral on social media.

Andrew Ng wrote a letter on the moment. His framing: loops are "a key part of how we get AI agents to iterate at length to build software." The implicit contrast is with one-shot prompting — where the agent gets a single attempt and the developer reviews the output. A loop-first design flips the priority: structure the iteration cycle first, write the prompts that serve it.

Naming matters in engineering disciplines. When a practice graduates from "the thing we do" to "loop engineering," it accelerates — practitioners compare notes, share patterns, and build shared vocabulary for the failure modes. The questions that follow naming are the productive ones: how many iterations to budget per task, how to checkpoint progress inside a long-running agent, and how the loop knows it's finished rather than just stuck.

For teams building on AI agents today, this is less a warning and more a prompt. The tools exist, the name exists, and the discipline is being written right now. Teams that engineer their loops deliberately — rather than hoping the model figures it out — will have a structural advantage that compounds as agents take on more complex work.

via @AndrewYNg on X →
02 Key News

AI adopters' workers safer than those at firms sitting it out

New research reported by the Los Angeles Times inverts the standard AI-job-loss narrative: workers at companies actively embracing AI may be safer from displacement, not more at risk. The finding points toward laggard firms — those holding back on adoption — as the more exposed end of the workforce risk curve.

via Los Angeles Times →

Companies quietly rehire workers as AI automation disappoints

Companies that cut staff in favor of AI automation are rehiring workers, according to QZ — a signal that the gap between what automation was sold to do and what it can actually handle in production is wider than the initial headlines suggested.

via QZ →

Enterprise AI: real deployment or still a slide deck?

Futurum Group examines whether enterprise AI has genuinely crossed from experimentation into production work — or whether most organizations' AI strategies remain aspirational.

via The Futurum Group →
03 Learnings

Andrew Ng: all his tasks now run through AI agents

A free 31-minute walkthrough on self-improving agent architecture — from one of the field's most concrete teachers.

Andrew Ng said 100% of his tasks are now handled by AI agents, with self-improving loops as the next step. His timeline: prompting disappears as a relevant skill within three to six months.

The signal worth acting on is the 31-minute video linked in the tweet — a concrete walkthrough on building self-improving agents from scratch. For engineers deciding where to invest learning time right now, a dense free session from one of ML's clearest teachers is a better use of an afternoon than most paid courses on the topic.

100% of my tasks are done by ai agents, self-improving loops are next.
via @AnatoliKopadze on X →
04 Tools & Craft

Open-source skill files give AI agents a design taste layer

TasteSkill's SKILL.md files stop Cursor, Claude Code, and other coding agents from defaulting to generic-looking frontends.

TasteSkill ships open-source SKILL.md files that plug into Cursor, Claude Code, Codex, Gemini CLI, v0, Lovable, and more. The v2 default skill reads the brief before generating — inferring industry, audience, mood, motion depth, and layout family — then maps to a real design system rather than a template. A hard pre-flight checklist must pass before output ships. One install command, any compatible agent.

The agent reads the room before generating: industry, audience, mood, motion depth, layout family.
via Taste Skill →

AI coding agent repos, ranked by GitHub stars

Spotted on LinkedIn: a roundup putting the current star counts in one place. Coding agents: OpenClaw (278K), Opencode (118K), Claude Code (75K), Superpowers (73K), Codex (63K). Agent tooling: Firecrawl (89K), Context7 (48K), Scrapling (25K). Infrastructure: Open WebUI (126K), llama.cpp (97K), Daytona (63K). Star counts lag real usage, but they're an honest proxy for where community weight is settling.

via LinkedIn →
05 From the Timeline

@bcherny's loop engineering post became the week's anchor thread

Boris Cherny — creator of Claude Code — posted what became the week's most-discussed take on loop engineering: 17K likes, 795 replies. The discourse will bury the original fast; worth reading it first.

@bcherny on X →
The Last Word
Automation promised to replace the workers. Now it's asking them back.
The Desk Report

How this edition came together — from bookmarks and feeds to the page.

110links gathered
30read by the desk
8made the edition

Where they came from

On the cutting-room floor — 102 links read but not run this week

Quality over volume: most links get a second look and a pass. The ones that made it earned their place.

Reading Room — every Sunday

The week's AI signal in 4 minutes — what happened, why it matters, and what to do with it. Curated by someone who actually builds, not a feed algorithm.