social-stream · 2026-06-09

2026-06-09-evening

Summary

The sharpest signal this evening is a two-post cluster on parallel agent orchestration in coding tools: Anthropic's bcherny shipped nested subagent support in Claude Code (agents spawning agents, capped at depth=5), and Kilo Code shipped its "Agent Manager" that runs multiple agents in isolated git worktrees side by side. Both land the same week, both bet that managing context across many concurrent agents is the next coding-tool frontier. The second real story is state money flooding into AI: ns123abc reports China is committing $295B over five years to build an 80%-domestic AI data-center network by 2028, while Scobleizer's newsletter argues governments broadly now want equity stakes in AI firms. Standout one-offs: eliebakouch flags a leaked Claude Mythos (Fable) price of $10/$50 per MTok, undercutting rumored GPT 5.5 pro by roughly 3x, and ns123abc pushes an OpenAI super-PAC story (Brockman's $25M donation). The bulk of the slot, brivael's 17 posts, is French political commentary and SpaceX/Musk hype with no AI substance, so it is noise.

Posts

  • Nested subagents in Claude Code (@bcherny). Anthropic shipped agents that kick off other agents to better manage context, capped at depth=5 in today's release. A concrete step toward recursive agent orchestration. See Claude Code architecture.
  • Kilo Agent Manager: parallel agents in git worktrees (cluster of 3) (@kilocode). Day 1 shipped multiple agents running in isolated worktrees with parallel execution; a follow-up demos two models doing the same task side by side so you keep the diff you prefer and discard the other. Same parallel-orchestration bet as bcherny's nested subagents, from an independent tool.
  • Nemotron 3 Ultra tops PinchBench open-weight (@kilocode). Kilo says NVIDIA's Nemotron 3 Ultra is the #1 open-weight model on PinchBench, their agentic coding benchmark, with a NVIDIA webinar on the architecture June 10. Webinar registration is promo, but the benchmark claim is worth tracking.
  • Claude Mythos (Fable) pricing leak (@eliebakouch). Citing The Information, a rumored $10 input / $50 output per MTok for Claude Mythos, roughly a third of rumored GPT 5.5 pro ($30/$180). If real, an aggressive price move on the next Claude tier.
  • China's $295B AI national project (@ns123abc). Claims a 2-trillion-yuan, five-year, debt-funded push for a unified national AI network by 2028, 80% on domestic chips from Huawei and others, state telecoms running the infrastructure. Big if accurate; treat the specific numbers as unverified social reporting.
  • Governments want a stake in AI companies (@Scobleizer · newsletter). Argues AI is becoming an investment race where the US debates public stakes while the UK and South Korea already invest directly in compute and chips. Pairs with the China item as the policy thread of the slot.
  • OpenAI super-PAC story (@ns123abc · The Hill). Reports Greg Brockman and his wife donated $25M to a "Leading the Future" PAC; OpenAI denies affiliation though its chief global affairs officer reportedly helped launch it. Click through for The Hill's reporting.
  • Tesla FSD Supervised approved in Denmark (@Tesla). Regulatory approval plus a V14.3.3 driving clip. Incremental rollout news, low relevance here.
  • brivael's French political and Musk-hype thread (cluster of 17) (@brivael). Sixteen of seventeen posts are French domestic politics, anti-socialism commentary, and SpaceX/Musk net-worth hype with no AI content. Skip.