Detailed writeups on multi-agent orchestration, AI engineering patterns, and what actually works in production.
After building several personal autonomous loops, I kept returning to the simplest design: treat Codex or Claude as a short-lived worker inside a Bash loop. The shell owns state, scheduling, validation, recovery, and done conditions. The agent does one bounded unit of work and returns JSON.
Pretext didn't introduce the loop to me. It reinforced the stricter version: lock the architecture, measure against reality, isolate the miss, classify it, and use AI for throughput instead of authority.
I spent two weeks benchmarking agent skills on Steel. The surprising result wasn't that prompts matter. It was that variance is expensive, browsing makes it obvious, and the biggest unlock came from redesigning the CLI so the environment carried more of the load.
If you want an AI-native dev team, don't start with autonomous coding demos. Start with requirement quality, design systems, task schemas, centralized memory, and explicit validation. AI amplifies structure. It also amplifies chaos.
The printing press took 60 years to become economically sustainable. LLMs are on a similar diffusion path—broad adoption, shallow integration. The winners won't have model access—they'll have the complements: context, workflow, trust.
The clean way to use Claude Code with both your normal account and z.ai is one neutral config and two simple entry points.
Copyright does not disappear in the AI era, but it stops functioning as a meaningful scarcity mechanism for software. As reimplementation gets cheap, the real moats shift to trust, governance, provenance, maintenance, and operational legitimacy.
Give the agent a Steel CLI and SKILL.md contract, force a snapshot/click/fill loop, then turn the successful run into a rerunnable bash script.
There's a hidden layer between human questions and search results. AI tools translate messy prompts into precise queries - and you can see the evidence in Google Search Console.
The best coding agents aren't about smarter models. They're about harness design: minimal core, extension hooks, radical transparency, real sandboxing, session forking, and headless RPC.
Request an AI summary of this blog
