In the 48 hours after OpenAI’s 7/9 dual launch (GPT-5.6 and ChatGPT Work on the same day), the agent ecosystem started realigning along three visible axes. OpenAI temporarily lifted the 5-hour rate limit on Plus / Business / Pro. Codex weekly active users crossed 6 million. And one production agent — Ploy, who builds real marketing websites with an agent that plans across apps, reads and writes code, takes its own screenshots, and decides for itself when the build is done — migrated its core model from Claude Opus 4.8 to GPT-5.6 Sol and posted first-impression numbers that beat Opus 4.8 across the board: 2.17x faster wall-clock, 27% lower cost, 0.970 vs 0.936 on the visual score. ...

On July 12, OpenAI quietly posted Prompting guidance for GPT-5.6 Sol on its developer site — a single, official recipe book for how to write prompts for the new model family. Read alongside this morning’s https://cuigh.com/posts/gpt-5-6-sol-production-migration-2026/ Ploy piece, it reads as the other half of the 7/9 launch story. The headline number is sharp: on OpenAI’s own internal coding-agent eval, swapping the system prompt from the GPT-5-era “encyclopedia” shape to a lean shape lifted eval scores 10–15%, cut total tokens 41–66%, and reduced cost 33–67%. That is not “the model is smarter so you can write shorter”. It is the reversal of prompt-design philosophy itself, now that model capability has caught up. ...

On July 9, OpenAI did two things on the same day: pushed GPT-5.6 to the top of the “strongest model” list, and shipped ChatGPT Work — an agent that can run across web, mobile, and desktop to complete hours of work for you. Sam Altman posted on X: “obviously the best model we have ever produced, but also one of the best blog posts we have ever produced.” That line looks like product PR. It is not. What Altman is actually saying is that neither the model nor the blog post is the decisive asset anymore. An agent that can run autonomously for hours is. ...

On July 4, a widely-circulated post by @AYi_AInotes highlighted a line buried in SpaceX’s revised IPO filings: SpaceX is supplying Anthropic with $1.25 billion of compute every month, under a contract running through May 2029, with either side able to terminate on 90 days notice. The same window saw reports that Anthropic has locked in 1.4 GW of data-center capacity in Australia, with $15 billion in build-out cost. Read these three together and the story is not “Anthropic bought more compute.” It is that AI compute has quietly moved from being a cloud resource you rent by the token, to a piece of industrial infrastructure you buy on multi-year fixed contracts. In industrial-age language this is called a power purchase agreement (PPA). In the AI era it is still the same shape: a customer locks in capacity at a fixed price, the supplier locks in cash flow to build. I am calling this “compute as power plant,” and that is the thesis of this post. ...

Reading my feed on July 2 and 3 — Hacker News, X, blogs, WeChat — five engineering tips connected themselves into a single line in my head. The line used to live scattered across sources without anyone naming it. When it surfaced in 24 hours, it became hard to ignore. I am not going to write an “agent tool roundup” today — I did that on June 28 with a five-layer stack. That was the map. This piece wants to be the signposts. The thesis is simple: five teams that actually use agents in production hit the same wall in early July. They each gave their own answer. I want to put those answers on the same page and see what is shared. ...

I need to start by correcting something I wrote on June 5. In Anthropic’s overnight triple move: building agents and containing them at the same time, I read Anthropic as a company running “two parallel lines” — a public “contain Claude” engineering program and a quieter recursive-self-improvement research track — and concluded that Anthropic was the safety-first operator, walking a “guardrails first, then capability” path. I had three pieces of evidence: an open-source vulnerability discovery framework, a recursive self-improvement research post, and the engineering article “How we contain Claude across products.” I read those three together and concluded Anthropic’s engineering culture was more credible than OpenAI’s. ...

I am writing this on the morning of June 30, 2026. The last day of June. If I had to summarize the most important engineering shift in AI in the first half of 2026 in one sentence, I would not say “the models got bigger” or “context windows got longer.” I would say: the protocols around agents are starting to grow up the same way the protocols around the web did in 1995. ...

AI agents are not one category anymore. That is the most confident thing I can say after a week of radar pulls. As of June 28, every agent-flavored project that consistently lands on the GitHub monthly leaderboard, the Product Hunt monthly leaderboard, or the HN weekly leaderboard has settled into one of five layers. Each layer has its own champion, its own growth curve, its own user base. The five: research, memory, orchestration, anti-slop, evaluation. This post lays the map flat, then explains why “separated cleanly” matters more than “built strong.” ...

On the afternoon of June 26 (U.S. Eastern), OpenAI published a preview post for GPT-5.6 Sol on its blog, and parked the system card at deploymentsafety.openai.com. The same day, the Washington Post reported that the U.S. government will decide which users get access to GPT-5.6. The same day, Reuters reported Anthropic is releasing Mythos only to “trusted partners.” Three events, less than 24 hours apart, each individually heavy. I am not going to write “AI regulation is here” — that ground was covered on June 14 in AI regulation just walked into the product. What I want to do today is zoom in much closer: what does GPT-5.6 Sol’s launch actually change, for a normal user, a normal builder, a normal product spec writer. ...