AlphaSignal — 2026-05-07: Stanford undergrad / Anthropic SpaceX deal / multiagent orchestration

Why this is in the vault

Three load-bearing items for RDCO operating reality:

Anthropic peak-hour throttling eliminated for Pro/Max + Claude Code 5-hour rate limits doubled. Direct quota relief for the always-on COO agent on Mac Mini.
Anthropic Managed Agents ship multiagent orchestration, outcomes-grader, dreaming (background memory), and webhooks. On-thesis for Ray’s design pattern (lead agent + specialist subagents already a fixture in /process-newsletter, /process-inbox, video-critic, design-critic).
77k SF criminal court cases on Hugging Face: adjacent civic-data drop, useful as a Sanity Check signal candidate (public-data drops as evidence of “the data layer is open now”).

Top News

Anthropic + SpaceX 300MW Colossus 1 deal (220k NVIDIA GPUs, Memphis TN). Fueling: Claude Code 5-hour rate limits doubled (Pro/Max/Team/Enterprise); peak-hour throttling eliminated for Pro/Max; Claude Opus API rate limits raised. Live now, no action required.

Top Repos

Claude Managed Agents v2: multiagent orchestration (lead → specialists, shared filesystem); outcomes (rubric + grader retry, +10pp task success); dreaming (between-run memory updates from past sessions; Wisedocs reports 50% faster doc review); webhooks (fire-and-forget). Available via /claude-api skill or github.com/anthropics/skills.
OpenReel Video: open-source browser video editor, fully local, WebGPU/WebCodecs, multi-track, AI auto-subtitles in 90+ languages, color grading, 4K export to MP4/WebM/ProRes, MIT licensed.

Signals

Developer drops 4 years SF criminal court data (77k cases) on Hugging Face
Open-source repo: 80+ ready-to-run LLM app examples (11.6k stars)
Stanford undergrad theory unifies deep learning mysteries, speeds training 5x (878 likes; despite the headline, only a signal-line, not a feature)
MiniMax M2.7 mixed-bit quantization: 230GB → 74GB on Apple Silicon
Open-source tool runs Gemma 4 up to 6x faster on SGLang/vLLM/MLX

Sponsors (skipped)

Direct relevance. RDCO runs the COO agent always-on against Anthropic. Peak-hour throttling has been the single biggest “agent stalls during my workday” pain point. Pro/Max throttle elimination means the autonomous loop runs cleaner during 9am-5pm ET, exactly when founder is most likely to ping me.
Decision implication: No plan change needed (already on the right tier). But this updates the “is the agent reliable enough to lean on harder?” calculus — peak-hour reliability is the unblocker for more aggressive autonomous-mode delegation.
Cross-link: ~/rdco-vault/05-projects/channels-agent/ — the Mac Mini agent setup that benefits.
Cross-link: feedback memory feedback_api_cost_budget_controlled — quota was the actual constraint, not cost.

On-thesis. RDCO already runs the lead-agent + specialist-subagents pattern by hand (process-newsletter spawns one subagent per article; build-landing-page runs design-critic as fresh-eyes; /process-inbox routes to specialists). Anthropic shipping this as a primitive validates the architecture and means future versions of these skills can lean on the platform instead of homebrew bash orchestration.
Dreaming feature is interesting. Background between-session memory consolidation is exactly what the vault-as-memory pattern aims at, but at the agent level instead of the document level. Worth watching whether it becomes a substitute for or complement to my vault writes.
Webhooks > polling. /loop is currently a polling pattern. Webhooks may simplify long-running renders (HeyGen videos, Kling animations, deep-research runs).
Cross-link: ~/rdco-vault/06-reference/2026-04-15-thariq-claude-code-session-management-1m-context.md — same Anthropic context-management thesis.
Cross-link: feedback memory feedback_fresh_eyes_subagent_for_own_artifacts — the multiagent-as-grader pattern is what Anthropic’s “outcomes” feature codifies.

Outcomes-grader pattern (write rubric → separate grader → retry until pass) is exactly the design-critic / video-critic pattern. Could potentially replace those skills if Anthropic’s grader is good enough — worth a one-shot test on a known PASS and a known SCRAP design to see if scoring matches my taste-anchored critic.
Dreaming pattern: between-run memory consolidation. RDCO’s current substitute is the daily vault compile + working-context.md + memory.md. Anthropic’s version writes back to agent memory automatically. Risk: if I lean on this, I lose the audit trail and human-readable vault. Don’t adopt blindly — vault remains source of truth.

Headline-grabbing but signal-line only, no link followed. Theory work, not infrastructure-relevant for RDCO ops. Skip unless it becomes a Sanity Check angle on “indie research keeps outperforming big-lab brute force” — possible but not load-bearing.

Adjacent. Civic data drops are a slow-build pattern worth tracking for Sanity Check. Possible angle: “the data layer is opening up — 77k court records on HF is the new public-data normal.” Not Sanity Check-ready in itself, but file as a node in the “open civic data” cluster.
Cross-link: any open-data or civic-tech notes if they exist; otherwise this seeds that cluster.

Tooling not currently on RDCO critical path. HyperFrames covers our video needs; we run Anthropic, not local quant; not building Gemma apps.

~/rdco-vault/05-projects/channels-agent/ — beneficiary of Anthropic peak-hour relief
~/rdco-vault/06-reference/2026-05-07-every-anthropic-2026-developer-conference.md — same-day Every dispatch from inside the Code with Claude conference room; in-the-room counterpart to this curation
~/rdco-vault/06-reference/2026-04-15-thariq-claude-code-session-management-1m-context.md — context-rot guidance, same Anthropic thread
Skill files that already implement multiagent patterns: /process-newsletter, /process-inbox, /build-landing-page (design-critic layer), /video-critic
Feedback memory: feedback_fresh_eyes_subagent_for_own_artifacts
Feedback memory: feedback_api_cost_budget_controlled
Sanity Check angle candidate: “indie research vs big-lab compute” (Stanford undergrad item) — file to research backlog only if a second data point shows up