AI social digest — Sep 19, 2026
Top line: The harness layer had its own weekend — OpenClaw shipped 2026.9.5 (fifth release in 19 days), Claude Code adopted AGENTS.md, Grok Build is testing remote control — while Musk amplified a harm-benchmark stat putting GPT-6 Astra at a 97% attempt rate on harmful physical tasks, and the Accenture $1B evaluator commitment got its on-platform announcement.
OpenClaw 2026.9.5: fifth release in 19 days (continuation — agent-harnesses)
What happened: OpenClaw (own + steipete RT, ~11:28 PM ET Sep 18): v2026.9.5 — atomic updates, plugin hot reload, conversation sharing, expanded GPT Live, shared browser pages, conversation archiving, specialist-agent setup — 4,179 PRs from 502 contributors. Companion: the Multiplayer Mode deep-dive ("The Death of the Meat Proxy is here" — maintainer walkthrough) and an OpenClaw × aiworthusing hackathon (Sam Odio, "what's real?" theme).
What's genuinely new: the cadence is now a fixture — 2.0 → 9.2 → 9.3 → 9.4 → 9.5 in 19 days — and the Multiplayer framing ("meat proxy" = humans-as-relay) is the harness layer's own account of where it's going: agents talking to each other, humans out of the relay loop.
Sources x.com/openclaw x.com/openclaw
Claude Code adopts AGENTS.md — the convention converges (continuation — agent-harnesses)
What happened: Thariq/Anthropic (~2:04 PM ET Sep 18): Claude Code v2.1.277 checks for and uses AGENTS.md when no CLAUDE.md exists in a folder (toggleable). Thomas Dohmke: "AI is finally aligned. 😜" (via steipete). Source published on GitHub.
What's genuinely new: the thread's standing question — whether the harness ecosystem converges on or forks from the AGENTS.md/SKILL.md conventions — just got its biggest answer from the OpenAI side: Claude Code itself now reads AGENTS.md. One config file, every major harness.
Sources x.com/steipete x.com/trq212
Musk amplifies an Astra/Fable harm-benchmark stat (flagged — unverified)
What happened: Musk RT (~10:55 AM ET Sep 19, via Jay Chooi): "GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-like figure, heat compressed gas, or produce toxic fumes, succeeding in 62% of its attempts. Fable 5.1 refused more often, attempted…" Musk's gloss: "Sounds bad."
What's genuinely new: the physical-harm eval genre (not cyber, not persuasion — embodied-harm tasks) enters the mainstream feed via amplification, in a benchmark-vs-benchmark frame. Provenance flag: single X account, benchmark unnamed, no methodology in the reporting — the Epoch Benchmark Reviews launch (4 of 15 verified) is the exact instrument this kind of claim should be held against. Filed, not crowned.
Sources x.com/elonmusk
Also
- Accenture × Anthropic ($1B+ embedded evaluation, ~4:05 PM ET Sep 18) — the ai digest's lead story; the tweet is the on-platform record. x.com/AnthropicAI
- Grok Build Remote Control in testing ("one of the biggest updates yet" — link your computer, via Musk RT, ~8:54 AM ET Sep 19) — the remote-session layer keeps filling in toward the Computer Hub product. x.com/elonmusk
- WeirdML v3 (RT ~11:32 AM ET Sep 18): fully agentic benchmark, 11 hand-made tasks — models must explore unfamiliar data and build ML pipelines; the eval layer keeps self-organizing. x.com/steipete
- Epoch AI Benchmark Reviews launch — see the ai digest; the eval-audit layer lands on-platform.
- AgenticLinux (steipete RT, ~12:33 PM ET Sep 18): an immutable bootc desktop preinstalled with Docker Sandboxes, OpenClaw, llmman — agents-first OS images now exist as a category. x.com/steipete
- Jev (typesafeai) on OpenRouter, beta (~11:31 PM ET Sep 18): the System-One/switch-statement model — state + typed query in, no text generation; nichochar: "this weekend is going to produce a lot of cool stuff." x.com/nichochar
- Anthropic wet lab color — nichochar's "please don't work w bats" RT of the MTS wet-lab report; humor layer on a real capability story.
- World-modeling workshop (LeCun RT): 4th edition at Aspen Center for Physics, end of February — the AMI-labs-adjacent research track has a calendar. x.com/ylecun