Agents at work ①
- 01 Plan–Code–Execute: agents that create their own tools
- 02 Vercel's Agent Skills: the missing middle layer
- 03 The next bottleneck isn't model intelligence
- 04 Give Claude Code a map (Graphify)
4 notes · ~30 min
Fresh from the workbench
I Built a Spec-First Framework for Coding Agents. It Made Them Worse.
A postmortem of NightLoop: why per-task scoped context degrades frontier coding agents, and why governance should verify outcomes instead of constraining what the model sees.
Read the note →Reading routes
All routes → The Dark Factory files ②
- 01 Why AI writes the code and devs are still slower
- 02 Inside the Dark Factory
- 03 The web is forking in the age of agents
3 notes · ~22 min
Curiosity cabinet ③
- 01 Do JPEGs age?
- 02 Why Latin isn't called Roman
- 03 Only ⅓ of Universal Grammar rules survived testing
3 notes · ~18 min
Latest notes
Archive →I Built a Spec-First Framework for Coding Agents. It Made Them Worse.
A postmortem of NightLoop: why per-task scoped context degrades frontier coding agents, and why governance should verify outcomes instead of constraining what the model sees.
5 Jul, 2026 · 4 min →
When a Model Becomes a Munition: The Fable & Mythos Suspension
The US government ordered Anthropic to cut off Fable 5 and Mythos 5 access for all foreign nationals on national-security grounds. Strip the outrage and a real precedent remains: frontier models are being treated like export-controlled weapons — and we've run this experiment before.
13 Jun, 2026 · 3 min →
Malware That Prompt-Injects Your AI Code Scanner
A reported campaign hides payloads from AI-powered security scanners by injecting weapons-related text that trips the scanner's own safety refusals. Single-source so far — but the technique class is worth understanding now.
13 Jun, 2026 · 2 min →
AMD's Gorgon Point Closes the Mini-PC Gap on the M4 Pro
A developer-and-LLM benchmark of AMD's new Ryzen AI mini PC against Apple's M4 Pro Mac Mini lands at a more interesting place than 'who wins' — it's about price-per-capability.
13 Jun, 2026 · 2 min →
Dynamic Workflows: When the Agent Writes Its Own Orchestration
Claude Code can now write an orchestration script and fan a single prompt out to hundreds of parallel subagents. What it's for, and the one reason not to reach for it.
13 Jun, 2026 · 2 min →
The Real WWDC Question Isn't Siri — It's Where the Work Runs
Apple's on-device AI push reframed as a bet on the most expensive question in the industry: when agents work all day, does that compute happen on your device or in someone's data center?
13 Jun, 2026 · 2 min →
The Spec Is the Work: Karpathy's Case Against Just Prompting
Andrej Karpathy's argument that the bottleneck in AI-assisted building isn't the model — it's the spec you fail to write. Distilled from a walkthrough of his AISN 2026 talk.
13 Jun, 2026 · 2 min →
Setting Up OpenAI's Codex: From Install to Colleague
A working path through Codex's five surfaces — install, approval modes, AGENTS.md, and the handoff protocol — distilled from Nate B. Jones' living guide, with the sharp edges marked.
12 Jun, 2026 · 5 min →
01
I Put Claude Code Inside Obsidian — and It Changed How I Use My Notes
02Plan–Code–Execute: Building Agents That Create Their Own Tools
03Vercel's Agent Skills: The Missing Middle Layer in Agent Architectures
04Do JPEGs age?
Fragments worth keeping, delivered occasionally.
No schedule, no spam — just the notes and experiments that survived the notebook.