Why this is in the vault
Anthropic merged Claude's chat and background-work (Cowork) modes and shipped Docs, Slides, and Design as native in-chat tools — directly relevant to how RDCO produces branded artifacts and whether external app-switching (Canva, Google Docs) stays necessary.
Mapping against Ray Data Co
RDCO already runs a Canva MCP connector and Google Docs/Slides workflows for lead magnets, decks, and design assets per ~/rdco-vault/CAPABILITIES.md; this collapses those into native Claude tools (Claude Docs for write/comment, Claude Slides for deck-build-and-present, Claude Design for inline visual mockups) inside one chat, no copy-paste between apps. Worth testing against the existing 12-stage production workflow (feedback_ic_vs_production_mode) to see whether Claude Slides can replace the current pptx-skill-plus-manual-design loop for public-facing decks, and whether Claude Design is a viable first pass ahead of the branded-design skills in CAPABILITIES.md rather than a competitor to them. The chat/Cowork merge also matters operationally: it's the same "drop a task, close the laptop, it keeps working and asks when blocked" shape RDCO already runs via Workflow/Task dispatch (feedback_workflow_agent_output_integrity), now productized by Anthropic directly rather than hand-rolled. Rolling out in beta on paid plans now, Pro/Max soon — worth a hands-on test before assuming it changes the design-skill stack.
Curation section
- Top Paper — Google DeepMind Dream-RSI: a self-improving search agent that replays past attempts instead of running new expensive experiments, cutting AI search/inference calls up to 162x in some settings without retraining the underlying model. Relevant as a pattern (cheap replay-based optimization layered on top of an existing agent) more than a direct RDCO tool need right now.
- Top Model — LongCat-Video-Avatar 1.5: open-source (MIT license), turns a photo + audio clip into a talking-head video in ~1 minute for a 10-second clip, multi-person scenes supported. Commercially usable; a candidate for cheap avatar/presenter content if RDCO ever wants that format, but no active use case today.
- Signals: Qwen-2.5 JSON inference sped up 5x with no retraining; Google Cloud multi-agent demo on G4 VMs; Max Planck Institute model trained on pre-1930 text to strip modern historical bias; Google's WikiSkill converts agent experience into transferable reusable skills; distill-then-RL training recipe for reasoning models; OpenAI publishes a public framework for disclosing model misbehavior (the "OpenAI admits models hide mistakes" framing in the issue's intro).
⚠️ Sponsorship
Three sponsors ran in this issue: Tiger Data (TimescaleDB — Postgres extension for metrics/embeddings/analytics, $1000 trial credit offer), Prior Labs (TabPFN-3.5 tabular prediction model), and Google Cloud (multi-agent GPU demo, presented placement in Signals #2). Tiger Data and Google Cloud are both confirmed members of AlphaSignal's rotating sponsor pool from prior issues. Prior Labs is a new sponsor not previously seen in the tracked rotation. All three are curation-embedded native ads; standard AlphaSignal bias — vendor claims (162x, 5x, #1 benchmark rankings) are self-reported by the sponsors or the labs being covered and unverified here.
Related
- [[2026-09-18-alphasignal-claude-code-parallel-cloud-threads]]
- [[2026-09-16-alphasignal-salesforce-claude-crm-odyssey3-robot-brain]]
- [[feedback_ic_vs_production_mode]]
- [[feedback_workflow_agent_output_integrity]]