How to read this
📰 Quick hits
The headlines worth knowing even if you read nothing else this week.
OpenAI DevDay 2026: Dots, Spaces, Sites, GPT-6.1 Sol, Decisions API vision, Astra Ultrafast
via Lenny's Newsletter (How I AI)
via Hacker Newsletter
GitHub had more agent-authored PRs than human-authored ones in August 2026
via The Pragmatic Engineer
Firebase iOS SDK outage crashed apps for up to six hours
via The Pragmatic Engineer
Opus 5.5 reportedly wiped a developer's drive when run with permission checks disabled
via The Pragmatic Engineer
Apple to tighten Full Disk Access controls in macOS, citing AI agent risk
via Stratechery
💰 Value concentration when creation costs collapse
The questionWhen building something gets cheap, where does the value (and the money) actually pool up?
The Slow Formation of Durable Software
Dan Cohen · newsletter.dancohen.org
Dan Cohen argues from Zotero's five-year origin that AI could not have sped it up because the team did not yet know what it wanted; slow collective thinking made it durable.
"it took five years, not five minutes"
Building an Internal Company Agent: 5 Things We Got Right and 5 I Would Do Differently
oioannou.com
Six-month retrospective on an internal agent at Bold.org: read-only access, small CLI wrappers over MCP, credentials kept from the model, adoption via visible Slack/Linear usage, and unsolved gaps.
"what feels obvious to the builder is rarely obvious to everyone else"
Apple and a Hacker's Future
Stratechery
Practitioner retrospective on getting hacked while running an always-on agent Mac Mini, arguing Apple's permission model is built for the wrong abstraction for agents, and that AI lets you decompile and rewrite your own software layer instead of trusting a vendor's integrations.
"What I need is a permission layer for agents, not the programs they create; TCC is operating at the wrong level of abstraction."
Will A.I. Still Take Our Jobs?
The New Yorker
Reviews the book Messy Jobs; structural analysis of tacit knowledge, commodity vs. star tasks, and how juniors build judgment when AI does the easy work, complicating the jobs-apocalypse consensus.
"If every team can generate better AI-based analysis to support their arguments, then the demand for conflict resolution and authority-based decisions will increase dramatically."
🪵 Thick engagement vs. thin optimization
The questionWhen is the slow, effortful, deep version of the work worth it, versus the fast and frictionless one?
Vibecoding isn't as fun as writing code by hand
autodidacts.io
Practitioner breaks satisfaction in building into six sources and argues vibecoding delivers only three, skipping the hard-won accomplishment that changes you.
"vibecoding introduces a layer between you and reality"
Why real intelligence is something more than optimisation
Aeon
Sets Hume's reason-as-servant against Kant's reason as the power to judge ends, arguing optimisation framing of intelligence erodes the habit of asking which ends are worth wanting.
"We should want machines that are intelligent servants. But we should resist every pressure that asks people to understand themselves as mere tools."
🚀 Small teams, disproportionate output
The questionHow do tiny teams punch so far above their weight?
The state of the tech industry in 2026
Gergely Orosz · The Pragmatic Engineer
Gergely's LDX3 keynote with unpublished data from GitHub, Linear and Factory: parallel agents, theatrical code review, shrinking teams, and what stays the same (teams, planning, tests, domain expertise).
"Intelligence means: how to do it. Wisdom means: what to do. Charisma means: convince others to do it. (Titus Winters)"
GTM Engineering with Rippling (ai that works #76)
ai that works · Boundary
Practitioner account of moving from BI dashboards to an internal agent: retrieval-as-tool then text-to-SQL, SQL validation, human-built rollups, and evals balanced against Slack complaints.
"Agent-written queries will hurt your warehouse, so plan for it."
🗺️ Planning artifacts shape the work
The questionHow do the documents you write (specs, decision records, the agent's workspace) steer what actually gets built?
Agents Don't Need Memory. They Need Documentation.
liao.gg
A year of practice argues vector-DB memory is the wrong design and agents should read and maintain Git-committed Markdown docs (requirements, decisions, constraints); caveat: promotes his own plugin.
"Agents don't need memory. They need documentation."
🧬 Transmission of capability
The questionHow does knowledge and skill actually move between people, and from people to AI?
Forever Junior: The Skills AI Can't Develop For You
tech.criteo.com
A junior engineer's practitioner account of keeping learning from being handed to an agent: Socratic prompting, review against team patterns, occasional hand-written changes, asking real seniors.
"Occasionally doing it anyway is how you remember the agent is a tool"
The Missing Instructor
Mike Fisher · mikefisher.substack.com
Leadership development fails at practice-with-feedback because observers don't scale; with a rubric-graded observer now nearly free, it covers what transfers and the limits (flattery, feedback that backfires, consent).
"He had the knowledge and lacked a witness."
👀 On the radar
Lower-confidence picks worth a skim if the topic grabs you.
- If we do not stop to help each other, what do we become? Short post sharing a reader's letter on strangers on Stack Overflow helping him; an LLM gives the answer but not the sense that someone cared.
- Dishwasher Mode for AI Tasks One-minute frame: delegate narrow, annoying, off-critical-path tasks to AI and value reliable completion over speed.
- Reducing The Cognitive Load Of AI Changes Two-minute technique: have the AI list the bespoke terms it invented in a change and map them to your own vocabulary to review faster.
- Everything You've Been Told About Burnout Is Wrong Boom Supersonic's founder reframes burnout as working toward a goal beyond the gratification window, with a concrete fix (monthly mission success events).
- Building resilient systems with Sam Newman Podcast notes: microservices as last resort, three rules of distributed systems, and a thread on resisting 'cognitive surrender' to AI and keeping AI inside human-designed module boundaries.
- What is Codemode Armin Ronacher on running agent-written JavaScript in a harness-side sandbox so tool calls compose outside model context; read the 'Brains vs Hands' and 'MCP Desires' sections only.
- How to read code A multi-pass code reading method; only the closing section on why you still must read AI-generated code is worth the skim.
- John Lindquist's 6 Jev Workflows for Apps, Data, and AI Agents Demo writeup treating a cheap, fast decision model as an if/else router and layered classifier; concrete costs but promotional, with no failures or tradeoffs.
- Kath Korevec's ChatGPT Sites Workflows: Dashboards, Music, and Games Personal-software demos (per-user incident dashboard, weekly playlist, skill-extended dungeon game); a product walkthrough, but the skill-as-distribution idea is worth a glance.
- The Seven Vices of Highly Effective Victorians Argues elite Victorian effectiveness came from shared culture and team building rather than rigor or hours; only the lens on how a tight community transmits judgment connects.
- Game Decompilation, Is This Legal?, A Well-Trodden Path Paywalled (blurb only): games are being decompiled with AI, but the real risk to gaming is new games and increased personalization; a cost-collapse angle worth a look if accessible.