Detailed writeups on multi-agent orchestration, AI engineering patterns, and what actually works in production.
I had Claude Code pry its own deep-research workflow out of its binary, then pointed that workflow at a question about itself. The verdict: it searches wide and never doubles back. The systems it resembles do. The whole game is that second hop.
Anthropic named the three ways long agent runs fail: agentic laziness, self-preferential bias, goal drift. Building Wire, a browser agent on Steel, I'd fought all three without the words. What the post nails, and three things the browser substrate forces you to add.
Picking a model feels like an architecture decision. It mostly isn't. Swap a model under a fixed harness and your numbers wiggle; rebuild the harness under a fixed model and they move a lot. The only benchmark that gets a vote is your own workflow. A leaderboard orients you; it can't tell you what survives your production loop.
Same source, two channels. One middleware file parses Accept headers, falls back to UA sniffing for known bots, and respects sec-fetch-dest as a browser safeguard. HTML keeps a Link rel=alternate header so agents can discover the markdown variant. Plus llms.txt and llms-full.txt as prerendered static files. About 500 lines total.
The next evolution of extensibility isn't a better plugin system. It's agent-readable recipes that teach coding agents how to adapt external capabilities into your app's contracts. The extension is no longer just code — it's an installation procedure plus judgment.
Fine-tuned Qwen3 on Clojure. 30B SFT hits 83.8% best-of-16, smashing GPT-5.4's 64%. But RLVR with shaped rewards actually lowered the ceiling—the verifier loop matters more than the training method. Built a deployable agent from it. Data quality was the bottleneck all along.
AI agents can write code, but they can't prove it's correct, maintain architectural coherence, hold institutional context, or make value-laden decisions. These four problems are what stand between today's demos and tomorrow's production systems.
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.
Request an AI summary of this blog
