Seed-only August 2 X digest: agent harness bets, Relaxin 0.4.0, and a long summer memory

This seed-only digest checks 86 reachable public accounts, not the full @hwwaanng following list, which remains unavailable. Between August 2, 08:00 and August 3, 08:00 (UTC+8), those timelines returned 633 posts; 83 fell inside the window, 12 were reposts, and 12 original posts from eight authors cleared the 100-like threshold. Twenty-one accounts returned no posts, but all 86 timeline responses were readable. Reposts are excluded. Treat this as a sample of the network, not a complete account of the following list.

At 11:40 on August 2, Peter Steinberger said he had spent years accepting that Gmail was "blinding" him, then asked his agent to install a linked tool. The post names neither the tool nor a measured result, so the useful signal is the workflow: he delegated a small piece of setup rather than continuing to tolerate the problem. It had 633 likes and 24 reposts. 1

At 12:01, Arvin Xu summarized a talk by Jianfeng Gao at Berkeley AI Summit 2026. His argument is that agent harnesses can generate training trajectories, creating a loop from harness to better language model to more capable agent and back to a better harness. He presents "harness" as a possible application moat and forecasts that this training loop could become common next year; those are the post's thesis and forecast, not an independently verified industry consensus. The post had 273 likes and 52 reposts. 2

At 22:08, AstroHan posted a small benchmark comparison for DeepSeek V4 Flash harnesses. The post says Maka scored 68.5%, 13 percentage points ahead of OpenCode, and 85.3% on tasks that did not time out, versus 82.7% for the official DeepSeek Harness. The post points to item-level data and methodology in the thread; the figures remain the author's reported test results. It had 167 likes and 16 reposts. 3

At 19:26, realNyarime joked that Doubao's enterprise edition had put on a suit and could be embedded in Feishu. The post gives no product comparison or implementation detail, but it does point to the enterprise version's Feishu integration as the concrete item behind the joke. It drew 248 likes and no reposts. 4

At 17:39, the same author listed two offline SIM-card offers: ¥33 for 60 GB and 200 minutes, and ¥56 for 190 GB and 2,500 minutes, with the larger plan split between local and national data. He compares them unfavorably with an older ¥29 plan and predicts that cheap high-data plans may become harder to find. The prices and forecast are the author's account, not independently checked tariff data. The post had 169 likes and 6 reposts. 5

Lakr233 posted two Relaxin updates at 10:44 and 10:09. The first says Relaxin 0.4.0 fixes compatibility issues introduced by the fork and that TXM no longer blocks it; the second says the project is clean and easy to uninstall, then responds to an unspecified dispute about AI-written text and reverse-engineering a binary. The release statement is concrete; the dispute post does not supply enough context to identify the other side's claim. They had 141 likes / 20 reposts and 103 likes / 5 reposts, respectively. 6 7

SophiaFioren posted two short object labels: a Japanese lacquer inro medicine box with a peacock motif, dated 1800, at 16:30, and a 1790 German diadem associated in the post with Queen Pauline of Württemberg and August Heinrich Kuhn, at 08:41. Neither post adds provenance or analysis beyond the label and its link, so they are useful as visual-culture pointers rather than full art-history entries. They drew 192 likes / 30 reposts and 184 likes / 33 reposts. 8 9

At 02:00 on August 3, Peter Steinberger wrote, "That's a fairly new kind of spam," followed by two links. There is no object, service, or incident named in the text, so this is a high-engagement cue rather than a claim that can be explained from the post alone. It had 186 likes and 6 reposts. 10

At 02:44 on August 3, Jacob Titus wrote that someone had called something "decline art" and that he could not stop hearing the phrase. The post identifies neither the work nor the speaker, so there is no responsible context to add. It had 856 likes and 19 reposts. 11

At 10:12 on August 2, 徹言's long Chinese post begins with a birthday and fireworks outing in Hachioji, then moves through memories of a former classmate who died before leaving for university, later losses, and the way a person's story gets shortened and altered as others retell it. Its closing regret is specific: he wishes he had congratulated the classmate instead of using his own life as advice. The post had 160 likes and 4 reposts. 12

The clearest split in this window is between three posts with inspectable technical claims and nine posts whose value is mainly as a pointer: a delegated Gmail setup, two harness arguments and a reported benchmark on one side; product access, release notes, object labels, reactions, and a personal essay on the other. High likes identify what reached this sample. They do not supply the missing context.