posts · 2026-07-24

This piece connects OpenAI's prompt guidance from GPT-4.1 through GPT-5.6 into a clear engineering shift: from hard-coding eight-step workflows into system prompts, to loosening control by scene, to GPT-5.6 outright recommending ablation tests to strip unnecessary steps. The logic is that models can now plan their own execution paths, so developers should move energy from "how to walk" to "what done looks like, how to prove it, and what the agent can't touch." The author validated this in a long-form translation system—ditching chunking and retry logic, replacing it with deliverable specs: no Chinese characters left, Markdown intact, terminology consistent. The agent decides whether to read the full text first or write a regex check script. I'd discount this a bit because the post doesn't share specific ablation metrics or before/after translation quality numbers. But the direction tracks: GPT-5's guide already mapped a spectrum from full model autonomy to strict code-controlled branching. GPT-5.6 just packaged that into an actionable method. If you maintain agent workflows, the practical takeaway is: don't trim prompts by feel. Run ablation tests on your eval set, remove one section at a time, and only keep what metrics say you need.

This one flips the usual instinct—waiting for a perfect model before shipping. The US Department of War's 2026 AI memo argues that the delay of waiting outweighs the risk of imperfect alignment, so new models must be deployable within 30 days of public release. Don't read this as "the military is handing kill decisions to AI." The post is clear: deployment doesn't mean surrendering final authority. Safety comes from four layers—constrain action boundaries first (read-only, sandbox), pause for human confirmation at critical decision points, verify behavior with execution receipts instead of self-reports, and make authorization dynamic with fast rollback. I'd discount this a bit because the post names zero specific models or tasks where this was tested, and there are no performance numbers. It reads more like a deployment principles checklist than a field report. If you're building Agent products, the two immediately useful bits are "no receipt, no trust" and "make rollback cheap instead of chasing perfection."