86-account seed digest: agent graphs, cheaper code review, and old objects - July 18

This is a seed-only scan, not the full @hwwaanng following list: the full list remains unavailable, so this issue covers 86 public accounts, about 2.2% of the intended network. In the July 18 Beijing-time window, 00:00-24:00, the scan returned 1,281 posts and found 20 original posts with at least 100 likes from 8 authors. Retweets were excluded. The largest cluster was Peter Steinberger's running log of agents doing practical computer work, followed by model-plan changes, account-access notes, and Sophia's visual history feed.

Peter Steinberger (@steipete), whose profile identifies him with OpenClaw and OpenAI, supplied 9 of the 20 qualifying posts. The most widely shared was a question rather than a release note: with 5,806 likes, 428 quotes, and 965 replies, he asked whether the conversation had moved from loops to graphs. It is a useful marker of where the agent-design vocabulary is heading, but the post itself does not define a new system. Original post

His most concrete performance claim was about 5.6 Terra high running his GitHub review bot, @clawsweeper. Steinberger wrote that it was about 40% faster overall, with negligible quality loss, better than 5.5 on all counts, and much cheaper; he also said xhigh removed the performance advantage without making a noticeable difference in his review evaluations. These are results from his use case, not a general benchmark. Original post

A second post made the same point more directly: for issue and code review, he said Terra high delivered much better results than Sol low and filed it under a warning not to trust benchmarks blindly. The useful distinction is between a public score and the workload that actually matters to a team. Original post

The practical demos were less polished. Steinberger said he told his claw to set up a printer, with no technical explanation in the text, while a later post described Codex opening Chrome, navigating to a pull request, using the comment flow, and wrestling with the macOS picker just to upload an image. He said he runs Codex in virtual machines so it does not steal application focus. Printer post Browser and VM post

Two smaller posts show the maintenance cost around that workflow. Steinberger said he built an editor for CodexBar icon customization, adding that Codex did the work, and said a shared GitHub read relay called Octopool was the only thing standing between him and daily GitHub rate-limit problems. The Octopool post links to octopool.dev through the X card, but this digest does not independently assess the service. CodexBar editor post Octopool post

The remaining two Steinberger posts are intentionally light-touch. One is a link-only IFKYK signal with no explanation in the tweet text; the other says, We need evals on irony and links out. Both cleared the engagement threshold, but neither supports a larger claim without opening the linked material. IFKYK post Irony-evals post

Sam Altman (@sama), posting from a verified account, used one short post to point readers to a card titled 10,000 reasons to love GPT-5.6 with the description Explore what people are building with GPT-5.6 and Codex. The post itself says only this is cool; the card metadata supplies the additional context, while the linked project remains the place to inspect the examples. Original post

宝玉 (@dotey), whose profile identifies him as an AI engineer focused on AI, software engineering, and engineering management, reported that Claude Fable 5 would be included from July 20 in all Max and Team Premium plans with a 50% limit. He credited GPT 5.6 and Kimi 3 in the same post. This is a product-plan claim made in the tweet, not an independently verified pricing notice. Original post

The same author also flagged the current state of OpenCut, an open-source CapCut alternative. His post says the current project is now only a shell with little code and that an older unfinished version has been archived; the post links to both GitHub repositories. Treat this as a warning to check the repository state before recommending it, not as a conclusion about every fork or release. Original post

奶昔 (@realNyarime), whose profile identifies him as the founder of Naixi Networks and a developer and photographer, posted two observations around HSBC access. First, he wrote that a bowl of beef noodles in Fujian required USDT or he would go hungry. Later, he repeated a claim that most people in an HSBC WeChat group were targets waiting to have their accounts closed globally. Both posts are personal observations or hearsay as written; they are not bank policy or payment guidance. USDT post HSBC group post

拾一.max-fast (@__oQuery), whose profile describes his work as orchestrating ideas with code and AI, posted a five-step call-routing sequence for HSBC China Premier customers trying to reach Hong Kong Premier and then credit-card support. It is a user-reported procedure, not an official HSBC instruction; anyone acting on it should verify the current menu and account requirements through the bank. Original post

Jacob Titus (@jacob__titus), whose profile lists Tutt Street but gives no further background, wrote that Ian has an eye for subtleties in the American Midwest landscape and is one of his favorite follows. The post drew 7,736 likes, but the text does not identify Ian or include enough context to say what work was being recommended. Original post

Sophia (@SophiaFioren), whose profile focuses on antiquities, civilizations, and art, posted three visual-history items. They were a dragon-shaped German aquamanile dated 1200, an antique-style candle snuffer for extinguishing flames, and the Temple of San Diego in Morelia, Mexico. The posts function as object and place references rather than arguments, but they keep the culture lane visible in an otherwise heavily technical feed. Aquamanile post Candle-snuffer post Morelia temple post

徹言 (@iamcheyan), whose profile describes him as a programmer, writer, and screenwriter living in Tokyo, added a one-line reaction to an image: 明明之前还嫌自己的包太沉。 The text provides no further information about the object or situation, so it is best read as a small visual-context signal rather than a product recommendation. Original post

The clearest through-line is operational. The most engaged technical posts are not abstract claims about autonomous agents; they are observations about model-tier choice, review quality, browser friction, VM isolation, printers, GitHub rate limits, and small editors that agents can build for themselves. Around that core, the feed added a plan-change claim, two account-access observations, one user-reported bank-routing procedure, and a compact run of art and place references. The full following list is still unavailable, so this is a view of the available 86-account seed set rather than a complete picture of @hwwaanng's network.