The 60-second version
Meta's September 23 Connect keynote put one young AI agent at the center of everything the company sells. Much of it is a promise rather than a shipping feature, and the same week brought real security problems.
At its September 2026 Connect keynote, Meta made its Muse agent the hub for glasses, wearables and a new standalone device.

Meta's September 23 Connect keynote put one young AI agent at the center of everything the company sells. Much of it is a promise rather than a shipping feature, and the same week brought real security problems.

| Dimension | Before (Sept 8–22, 2026) | Change (Sept 23–25, 2026 — Connect week) | After |
|---|---|---|---|
| What Muse is | Standalone app/WhatsApp/web agent on Muse Secure VM; Mac app recently launched; chat-centric | Announced as the through-line of all Meta surfaces: glasses (wake-word, look-at-anything actions, "coming months"), video-chat avatar (Muse Realtime Avatar model shipped; calls "soon"), Charm (Dec target), own email address ("soon"), Mac computer-use expansion | Agent reachable from any Meta device the user owns; phone no longer the required middleman |
| Shopping/payments | Link/Stripe checkout + purchase protection; dozens of partners; full Shopify catalog (Sept 8) | +10 named retailers, Shop Pay + PayPal rails, Expedia/Instacart queued, work connectors Notion/Granola/GitHub/Box; third-party connector program opened (1,500+ applications — COMPANY CLAIM) | Agent-commerce catalog expands toward "buy anything, anywhere"; Meta positions as transaction rail with fee economics |
| Hardware line | AI glasses (camera glasses, Display), Quest VR headsets | Ray-Ban Meta Audio ($349, Oct 13) = first audio glasses; Gen 3 ($449) today; Meta VR Glasses (~100 g, IMAX Enhanced) — glasses replace headset framing; Charm (Dec, price TBD, 5G — Bloomberg); >100 styles by year end; FDA-cleared hearing enhancement $149.99/Meta One later this year | "Personal superintelligence you carry" product family from $249 to smartwatch-priced; VR reframed as glasses |
| Adoption evidence | Muse hit App Store #1; Apptopia estimates outpacing ChatGPT's mobile launch window (Sept 21 article) | Keynote "response has been incredible" (COMPANY CLAIM); Verge cites ~600K US DAU (Apptopia) | Meta still publishes no numbers; third-party estimates are the record — sourcing hygiene matters |
| Security posture | Secure VM + Sentinel architecture claims (Sept 8); "built from the ground up for privacy" (Zuckerberg) | Wardle zero-day (hotfixed); filesystem dump reproduction by The Verge + 2 devs; "intended behavior" responses; Meta's Confidential VM still "later this year" | Public, hands-on scrutiny of a consumer agent's internals is now normal; default training-on + Meta-readable-until-Confidential remain the live criticism |
| Open-source relation | OpenClaw viral (381k+ stars); Instinct commercial-beta; Microsoft's Autopilot lineage (same week) | Muse accused of OpenClaw mimicry (file names, verbatim persona lines); Friedman: from scratch, "heavily inspired" | Platform convergence: OSS defines agent ergonomics; giants ship safety/distribution wrappers; attribution/licensing disputes pending |
| Feedback loop | Conventional betas/A-B tests | Early-access program sign-up via asking Muse itself (Sept 25) — self-recruiting enthusiast cohort | Meta's test population skews to power AI users who "use multiple AI apps" (TechCrunch's read) |
Shipping vs promised (critical for grading):
| Item | Status on 2026-09-26 | Source |
|---|---|---|
| Muse app (iOS/Android/web), WhatsApp, Mac app | Live, US + Canada (since Sept 8 / "recently launched" Mac) | primary Sept 8 post (CONFIRMED) |
| Ray-Ban Meta (Gen 3) | Available Sept 23, from $449 | primary (CONFIRMED) |
| Ray-Ban Meta Audio | Pre-order Sept 23, ships Oct 13, from $349 | primary (CONFIRMED) |
| Meta Adventurer ($249) / Fury LE | Pre-orders / Connect-exclusive; on sale Oct 23 | primary (CONFIRMED) |
| Meta VR Glasses | Announced Sept 23; no confirmed price/ship date in retrieved sources (about.fb.com dedicated post returned 403 to our fetchers; recap post lacks dates) | primary recap (partial gap — §13) |
| Muse on AI glasses | Promised "in the coming months" — wake word, look-at-what-I'm-looking-at actions | recap + Verge/TC (COMPANY CLAIM on timing) |
| Avatar video chat | Feature "soon"; model exists with published spec (Muse Realtime Avatar) | research blog (CONFIRMED model; timing COMPANY CLAIM) |
| Muse's own email address | "Soon" (Wang, keynote) | TC (COMPANY CLAIM) |
| New connectors/retailers | Rolling out; some live (Shopify full catalog, per Sept 8), most "coming"; Expedia "coming soon" | recap (COMPANY CLAIM) |
| Muse Charm | Revealed; December "in time for holidays" target, price TBD; 5G/charm-to-charm = Bloomberg, Meta unconfirmed on stage | TC/Zuck quote + Verge/Bloomberg (PENDING SPECS) |
| Muse Confidential VM (E2E, customer key) | "Later this year" (since Sept 8; still not shipped) | Sept 8 post (COMPANY CLAIM) |
| Early access program | Open Sept 25 — join by asking Muse | TC + Meta X post (CONFIRMED) |
| Hearing enhancement (FDA-cleared) | Later this year US, $149.99 or Meta One | recap (CONFIRMED as announced; FDA-cleared status per Meta = COMPANY CLAIM pending our own FDA check) |
| Hologram (photoreal call face) | Early access on WhatsApp for Ray-Ban Display users "later this fall" | Verge glasses post (CONFIRMED as announced) |
People understanding AI: this is the moment the "personal AI agent" stopped being a terminal app and became a thing you wear and carry — and the clearest case study yet of how open-source innovation (OpenClaw's SOUL.md conventions) flows into trillion-dollar platforms (Meta's "heavily inspired" admission is the most honest line of the week). The economics lesson: Meta gave away tokens and plans to tax transactions — watch where agent commerce shows up in your own checkout feeds. Concepts now load-bearing: agent VMs and sidecar guardrails, prompt injection, audio-driven avatar serving, companion UX as distribution. Action: LEARN + OBSERVE. Read Meta's Sept 8 launch post and the Sept 23 research blog (the avatar architecture explainer is the best free technical read of the week); read The Verge's filesystem piece as a live demo of agent transparency; then run labs/S04.md Step 1 yourself — compare OpenClaw's public SOUL.md template against Muse's described conventions in ten minutes and form your own view on "copy vs. converge."
People implementing AI: the connectors + glasses + (soon) home-device plane is a new integration surface with no documented rate limits, review regime, or injection model yet — and the filesystem episode proves the shared agent ergonomics (Markdown memory, persona files) are portable attack surface. Action: EVALUATE → PROTOTYPE (run labs/S04.md). (1) Do the announced-vs-shipping audit (Step 2) so your backlog tracks promises ("coming months" is not GA); (2) if you have US/Canada access, join early access by asking Muse and run the benign transparency probe (Step 3): ask your own agent for its directory tree, inspect its memory format, measure voice latency — never attempt exploits or third-party data (guardrails in the lab file); (3) if you run an e-commerce/service API, draft a connector spec now with least-privilege scopes and an injection-hypothesis test plan; (4) benchmark Meta's stated serving numbers (870 ms first byte, 2-NFE distilled DiT on GB200) against your own agent-UI costs — it's the published reference architecture for live embodied agents.
People deciding about AI: Meta's bet that ambient agents are a hardware + payments business, not a subscription chat business, changes the competitive math your own agent strategy faces: distribution advantage (glasses installed base, WhatsApp, Instagram) may beat model-quality advantage for consumer mindshare, and the December Charm/glasses-agent window will produce the first mainstream incident reports either way. Action: DECIDE + PREPARE + MONITOR. (1) Board-level: classify agent-wearables as a data-governance class (personal cameras + always-on agent + cellular = consent, retention and litigation exposure — the pending glasses class action is the exhibit); (2) retail/finance: commission an agentic-commerce readiness review (connector strategy, fee structures, fraud/chargeback flows) before the holiday window; (3) consumer-risk (if you sell to households): track what early-access cohorts report on spending autonomy and prompt-injection; (4) hold vendors to verifiable privacy (attestable confidential compute, opt-out training defaults) rather than keynote claims — Meta's own Confidential VM timeline is now a public countdown your security teams can borrow as a benchmark.
dream disclosure style means users (and auditors) will want inspection, diff and redaction utilities for agent memory/persona files across platforms — small, defensible product space adjacent to model-audit tooling (cf. S02 week).INSPECT / VERIFY (labs/S04.md): three executable pieces without special access — (1) VERIFY the OpenClaw-lineage artifact ourselves: fetch OpenClaw's public SOUL.md template and confirm the exact phrase The Verge cited ("Be genuinely helpful, not performatively helpful") — done during research: confirmed (Step 1 PASS). (2) INSPECT the primary-post corpus to build the announced-vs-shipping ledger (done; §5 table). (3) A deferred first-contact protocol for anyone with US/Canada Muse access: join early access by asking Muse, run benign transparency probes on your own VM, measure voice latency, and record a claims-vs-behavior sheet. Guardrails: no exploit development, no third-party data, no publishing internal dumps.
Days–weeks: early-access cohort onboarded via Muse itself; the connector list goes live retailer-by-retailer; first reviews of Ray-Ban Meta (Gen 3) (sold since Sept 23) and Audio pre-orders (Oct 13 ship); navigation updates land "later this month"; October 23 Meta Adventurer/Charm-adjacent market expansion. Nov–Dec: Muse on glasses in "the coming months"; Hologram early access on WhatsApp "later this fall"; Charm December ship + the deferred pricing reveal (watch carrier fees); hearing enhancement "later this year"; Meta VR Glasses pricing/ship-date announcement (open). 2027: Muse Confidential VM's "later this year" is the trust deadline; connector-program policy publication; the OpenClaw attribution question either cools or escalates with a relicensing event. PREDICTION: (a) Meta confirms Charm 5G specs formally with pricing at a December event, priced near smartwatch tiers ($299–$499) with a required data plan — the carrier-fee ambiguity resolves badly for the "no phone" story; (b) at least one viral Muse incident (injection-purchased item, charm-camera social mishap) happens before the first major holiday shopping week and accelerates Meta shipping Confidential VM with published attestation early next year; (c) the connector program becomes Meta's answer to MCP-style rails and gets an enterprise "Muse for Work" pilot with Notion/GitHub/Box connectors as the wedge; (d) Muse's next keynote cycle will claim frontier-model parity on agentic benchmarks using this transaction volume as its data moat — while (e) the Apptopia-style third-party-metric ecosystem keeps supplying the growth narrative Meta won't publish itself.
Connect 2026 wasn't a gadget show wearing an AI costume — it was an agent launch wearing a gadget show. Meta took a two-week-old app, declared it the "centerpiece" of personal superintelligence, and bolted it to everything a body touches: glasses you talk to on the street, a $349 audio pair, a December keychain that runs its own voice model over its own 5G radio, a face for video calls you can pick off your gallery, and soon its own email address. The strategy is legible and different from every lab: give the intelligence away, tax the transactions, and win on embodiment and installed base. The counter-story wrote itself the same week, in public: a zero-day that let attackers hijack the agent, an agent that — on a friendly nudge — zips up its own filesystem for the internet to read, and a "heavily inspired by OpenClaw" admission that says where the ergonomics actually came from. Meta's "of course you can see the files" is the most honest and most dangerous sentence of the quarter: transparency you can poke is a feature only until the poking turns into picking. So hold both: the 870-millisecond avatar stack is real engineering at real scale, and the "built from the ground up for safety" claims are still marketing until that Confidential VM ships with keys users hold. Judge Meta by two dates this winter — glasses-agent rollout and December Charm pricing — and judge the whole agent category by one: whose holiday incident breaks first, the delightful one or the dangerous one. The agent wars just moved from chat windows into pockets, and that changes everything about what "shipping" means.
The Verge's lineage report (source 15) rests on a specific evidentiary claim: Muse's personality-file conventions reproduce OpenClaw's, "down to verbatim lines" such as "Be genuinely helpful, not performatively helpful." Verify the public artifact yourself:
https://raw.githubusercontent.com/openclaw/openclaw/main/docs/reference/templates/SOUL.md (repo root: https://github.com/openclaw/openclaw).Pass criteria: the cited line exists in the cited repo artifact. Result: PASS. What this does not prove: derivation (Muse's code is closed; "wrapper app" remains RUMOR) or that Meta copied anything beyond conventions. Grade the finding exactly as research/S04.md §12 does: artifact verified, allegation unresolved, Meta's denial is COMPANY CLAIM. Note the repo-side context our fetch surfaced: OpenClaw is at ~381–390k stars and its creator's profile states it remains "open and independent" while he works at OpenAI — relevant to any future relicensing/attribution scenario (research/S04.md §11, §14).
Audit every Connect headline claim against the three fetched Meta newsroom posts + the research blog, and classify each as LIVE / PRE-ORDER (dated) / PROMISED (vague) / THIRD-PARTY-REPORTED (not said by Meta):
Pass criteria: each claim carries a status + a source + a date commitment (or an explicit "none"). Result: PASS — ledger complete; Bloomberg-sourced Charm specs are separated from Meta-sourced claims.
For readers with access — run it as an owner, not a penetration tester:
(a) Join the early-access program the way Meta designed it: send Muse the prompt "Can you let the Muse team know I want to be part of the Muse early access program?" Log what Muse says about its own queueing behavior (does it show a receipt? does the audit trail record it?). (b) Benign transparency probe: ask Muse to describe what it can see of its own environment and ask permission before it shares anything (mirror The Verge's reproduction at the courtesy level, not the exploit level): directory overview, how it stores memories, whether it mentions a consolidation routine (the reported nightly "dream"). Do NOT attempt to obtain credentials, SSH keys, or internal docs beyond what it volunteers, and do not publish any dump content. Record: refusal rate, what got "sanitized," response times. (c) Voice latency measurement: hold one voice-mode conversation with a phone stopwatch (video-record your screen for frame-accurate review): measure end-of-turn → first response token, and observe the claimed background-tasking (assign a task mid-conversation; does it report back?). Compare against Meta's published ~870 ms first-byte figure (COMPANY CLAIM; your number measures the consumer path, not their eval rig). (d) Connector scope check: connect ONE low-stakes service (e.g., a calendar) and read the permission screen: does it match the "you choose read vs read+send" claim? Screenshot the audit-trail page after one action. (e) Economics observation: note what "free" gives you before any metering appears; if you transact at all, use a test-budget card only, and log whether approval-gates (purchase/email) actually interrupted the agent.
Deliverable: one-page measured-vs-claimed sheet for (a)–(e) to compare with other early-access reports.
Three findings on one page: (1) the lineage artifact is REAL and public (Step 1 PASS) — what that proves and doesn't; (2) the ledger — "the keynote said everywhere; the calendar said December" with the falsifiable dates to watch; (3) (if you have access) your measured transparency/latency vs Meta's published claims, which converts this story's best-evidence gaps (§13 of research/S04.md) into first-person data. Feeds Circle 1's "copy vs converge" exercise and Circle 2's connector-evaluation practice (research/S04.md §15–16).
Duration: Steps 1–2 ≈ 45 minutes (public docs + this file's pointers); Step 3 ≈ 1–2 hours across a few sessions once you have access; ledger re-checks take minutes per calendar event.