Why this is in the vault
Claude's new Salesforce-CRM absorption, this issue's Ory-sponsored "give every AI agent its own identity" pitch, and a Signals item on LLMs silently absorbing personality traits from fictional training characters form a coherent agent-identity throughline that gives RDCO's live cross-substrate-agent-identity research a dated public data point, plus Odyssey-3 as a second independent confirmation of the "one backbone, many surfaces" consolidation pattern.
Mapping against Ray Data Co
Anthropic's Salesforce-in-Claude integration ships a permission model where "Claude cannot see anything a user could not already see" — permissions inherited directly from existing Salesforce roles, one admin connects it once, no separate provisioning step. That is the same inherited-permission shape RDCO already runs across Ray's own MCP integrations (Notion, Gmail, Google Calendar, Drive), the posture behind project_1password_agent_auth_posture and the feedback_mcp_install_security_review_default install gate. This issue is a public confirmation that the pattern RDCO already committed to is becoming the default enterprise-integration shape rather than a workaround unique to a small COO-agent setup. Second, directly load-bearing: the Ory sponsor block ("give every AI agent its own identity... preserves the delegation chain back to the person or process that started it") and the Signals study on LLMs absorbing personality traits from fictional characters both land on exactly the question [[2026-09-13-cross-substrate-agent-identity-registry-standard]] is actively tracking — whether agent identity becomes a portable, auditable primitive or stays locked inside whichever control plane issues it (Salesforce/Agentforce, Entra, GCP). Claude folding Salesforce's entire CRM into one chat interface is itself a data point for that brief's stated risk that "control-plane consolidation narrows the portability wedge" — one more surface just moved inside a vendor's own chat window instead of staying an externally-orchestrated agent call. Odyssey-3's "one shared frozen backbone + small task-specific adapters" is the same consolidation shape one level down the stack (world model, not CRM) — independent confirmation of the pattern this issue's own intro names outright ("one model, one interface, one backbone to rule them all").
Curation section
- Top News — Anthropic ships Claude's Salesforce integration with 37 built-in sales tools. "Salesforce in Claude," open beta. 37 pre-built skills (call prep, deal-health review, record updates, forecasts) run inside the chat window instead of the CRM UI; permissions inherit from existing Salesforce roles. Service/marketing/commerce skills planned later this year.
- Presented by Ory — "Give Every AI Agent Its Own Identity." Ory Agent Security gives every agent/sub-agent a distinct identity, checks permissions before tool calls, and preserves the delegation chain back to the originating person or process. Claims coverage across 11 harnesses and 14 SDK integrations. See Sponsorship.
- Top Repo — LongCat-Video-Avatar 1.5 (Meituan, MIT license). Open-source photo+audio → lip-synced talking-avatar video. Swaps in Whisper-Large for lip sync, supports two-character scenes and text-only character generation at 480p/720p. Self-hosting needs a 40GB GPU at ~44s compute per second of output; weights on Hugging Face.
- Presented by Flint AI — "We built a 50-person dev team. Only 5 are human." Switch: a shared room keeping 5 humans + 45 agents on the same session history across Slack/Teams/Discord. See Sponsorship.
- Top News — Odyssey-3, a universal world model for robots, cars, drones, and game agents. One frozen shared backbone plus lightweight task-specific adapters per embodiment. Generates interactive simulated worlds for training; claims 77% real-world driving-policy transfer from pure-simulation training on Indian roads. Public developer-portal access "within weeks."
- Signals (one line each): Google DeepMind ships Gemini 3.8 Live (real-time vision, 97-language voice); WorkOS Agent Registration lets agents self-register users via auth.md with scoped short-lived credentials (see Sponsorship); a study finds LLMs silently absorb personality traits from fictional training-data characters; a new RL method targets the "only helps on easy problems" failure mode; a mini-paper unifies goal-based and hierarchical RL via hidden Markov models; Sakana AI trains 1000-layer nets without backpropagation using local-only dynamics.
Zero deep-fetches this issue — the Salesforce integration, Odyssey-3, and the identity/personality items are all sufficiently specified in AlphaSignal's own summary for the mapping above; none of the third-party links offered a hook beyond what's already extracted.
⚠️ Sponsorship
Three "Presented by" blocks, all confirmed recurring pool members, no new entities today:
- Ory — Agent Security / agent identity pitch (Top News slot). Recurring — first seen 2026-09-09, recurred 2026-09-14, this is at least a third appearance.
- Flint AI — Switch, multi-agent shared-session tool (Top Repo slot). Recurring — first seen 2026-09-08.
- WorkOS — Agent Registration / auth.md scoped credentials (Signals #2). Recurring — first seen 2026-09-15, this is its second appearance.
Also present: the masthead "In Partnership with" logo slot, rendering with generic partner_image alt text and no resolvable name or link target even under FULL_CONTENT — consistent with the intermittent-resolvability pattern already logged (legible only once, 2026-09-09, as QA.tech).
None of this is disqualifying — treat every AlphaSignal issue as unknown-until-scanned per the standing README note; disclosed here so a later reader sees the angle on the Ory and WorkOS items in particular, both of which argue for their own category (agent identity/auth) inside an issue whose editorial frame is also about agent identity.
Related
[[2026-09-15-alphasignal-flybrain-robot-bytedance-self-improving-rlt]] [[2026-09-13-cross-substrate-agent-identity-registry-standard]] [[2026-03-19-not-boring-world-models]] [[project_l5_north_star_strategic_direction]] [[project_1password_agent_auth_posture]] [[feedback_mcp_install_security_review_default]]