AGENTV / 05-OPEN-QUESTIONS.md

Answer the questions that make the frame real.

Fourteen questions for VTV and hardware/product reviewers. P0 blocks the first shippable unit. P1 blocks scaling past early prototypes. P2 names the questions we want to get smarter about before the category hardens around us.

Back To PacketRaw MarkdownJSON Packet

P0

Blocks first ship

Q1 / P0

Is the Live Encounter primitive the right shape for an unsupervised always-on home unit?

AGENTV consumes the same chamber plus canvas primitive used by SOLIENNE and COLTRANE, but home mode may run for hours or days rather than a bounded gallery encounter.

What we want reviewed

  • Is direct_address_only the right wake posture, or does a home need silent-unless-addressed-and-human-present?
  • Does long-lived home presence need a new ambient phase, or should idle plus presenceMode cover it?
  • Can ignored noise still pollute participant.turn logs, and what transcript hygiene is required?
  • Do we need session.pause/session.resume for multi-hour home quiet periods?

References

planning/SPIRIT_LIVE_ENCOUNTER_PRIMITIVE_SPEC.md / lib/canvas/types.ts / lib/agents.ts

Q2 / P0

Is VAD-before-cloud defensible as a privacy posture for a European home?

The intended claim is that mic audio does not leave the device until voice activity is detected, with a visible indicator while the agent lane is listening.

What we want reviewed

  • Is browser-side VAD enough, or must enforcement be firmware/OS-level to be credible?
  • What consent regime applies in an EU private home, especially for guests?
  • Does emotion or relational tagging introduce EU AI Act risk?
  • How should the hardware prove the LED indicator is tied to mic state rather than being decorative?

References

AGENTV product privacy section / AGENTV hardware physical security section

Q3 / P0

Is the pairing UX defensible as both a security surface and a human ritual?

The primitive supports QR, short pair codes, direct display links, and Google Cast. The product question is which one should lead in a home.

What we want reviewed

  • Is a read-only 8-character pair code acceptable, or does /display need per-session auth?
  • Should the home ritual lead with Cast, QR, or short code?
  • If the canvas drops mid-encounter, should the agent notice and acknowledge it?
  • Are multiple canvases in one session a feature or a footgun?

References

app/chamber/pairing-card.tsx / app/cast/receiver/page.tsx / app/receiver/page.tsx

P1

Blocks scale

Q4 / P1

Is the compute split right for Home Mode?

Today the Pi is the presence endpoint, Vercel hosts the web/canvas layer, and a hosted RPC service handles cognition.

What we want reviewed

  • Should on-device transcription or more privacy-preserving inference move onto the Pi?
  • Is a single cognition host acceptable for early units, and at what fleet size does multi-region matter?
  • Does Paris-to-US latency change the product feel enough to require a European host?

References

AGENTV architecture packet / services/rpc/src/server.ts

Q5 / P1

What is the tunnel failure mode when WiFi drops?

The frame may lose collector WiFi, Vercel connectivity, or RPC tunnel connectivity while physically remaining on the wall.

What we want reviewed

  • Should the frame render a local offline/resting state from cached assets?
  • Does the phase vocabulary need offline, or is offline a transport diagnostic outside phase?
  • Should the agent speak from cached local audio after prolonged connection loss?
  • What is the self-recovery story when remote admin access disappears?

References

planning/DEVICE_CAPABILITY_ANNOUNCEMENT_2026-04-26.md / packages/canvas-kit/src/presence.ts

Q6 / P1

Will the three-secret auth model bite us?

Current data-plane secrets are fleet-level in the prototype. That is simple, but stolen-unit revocation becomes brittle.

What we want reviewed

  • When should secrets become per-unit rather than fleet-global?
  • Is unit-scoped Bearer auth sufficient for early hardware, or do we need OAuth/device enrollment?
  • What exactly happens when one unit is stolen or compromised?

References

app/api/session/route.ts / app/api/rpc/callback/route.ts / app/api/canvas/events/route.ts

Q7 / P1

What does cross-session memory look like years from now?

A home unit implies repeated long-term encounters. The memory system needs to age gracefully and remain inspectable.

What we want reviewed

  • Is a rolling summary enough, or do we need hierarchical episodic plus summary memory?
  • When does a collector memory summary become too dense to be useful?
  • Can collectors read and redact their own memory summaries?
  • How should redactions surface back into the live encounter and canvas?

References

lib/memory/providers.ts / lib/memory/memstore-writeback.ts

Q8 / P1

Does the commerce surface belong inside the primitive?

A voice-initiated offer might become a canvas-rendered checkout card, but the protocol currently treats this as synthesis rather than commerce.

What we want reviewed

  • Should CanvasArtifactKind gain commerce.offer and commerce.receipt?
  • Should pricing guardrails move into the live-canvas cognition layer?
  • What legal status does verbal assent plus QR confirmation have?
  • Who owns chargebacks, refunds, securities risk, and tax nexus?

References

lib/canvas/types.ts / services/rpc/src/visual-action.ts

P2

Get smarter

Q9 / P2

Wake word or VAD plus direct address?

A classical wake word is cleaner in a noisy home; direct address may feel more intimate.

What we want reviewed

  • Should SOLIENNE require a wake word in all home contexts?
  • Is direct address enough if VAD and attention policy already gate replies?

References

lib/attention.ts

Q10 / P2

Should idle canvas be static or generative ambient?

A live-updating canvas can feel inhabited, but also changes the product from quiet frame to daily artwork emitter.

What we want reviewed

  • Should idle show a slow-breathing static portrait or daily generated work?
  • Does daily generation require a separate consent or taste policy?

References

app/display/presence/page.tsx / packages/canvas-kit/src/presence.ts

Q11 / P2

How should one household handle multiple people?

The primitive can carry a speaker field, but the current brain largely treats a unit as one collector relationship.

What we want reviewed

  • Do we add diarization and per-person memory for spouses, kids, and guests?
  • Should v1 deliberately stay one primary collector, one relationship?

References

lib/canvas/types.ts

Q12 / P2

How should we design around the Runway long-session ceiling?

Long avatar sessions can drift away from backend cognition calls after repeated turns, which matters more at home than in short kiosks.

What we want reviewed

  • Should the product silently re-roll sessions every N turns?
  • Should a long home conversation explicitly close and reopen as a ritual?

References

lib/runway/session.ts / app/chamber/live-avatar-client.tsx

Q13 / P2

What happens on resale?

A collector may sell the unit. The memory and relationship boundary is philosophical as much as technical.

What we want reviewed

  • Should serial-number re-pair wipe memory by default?
  • Is memory exportable and re-inhabitable, or does the prior relationship end?

References

Spirit identity and memory discussions / lib/agent-manifest.ts

Q14 / P2

What should happen when the paired canvas device sleeps?

Phones sleep, TVs enter screensaver, and browser tabs get suspended. The agent lane may remain idle or active.

What we want reviewed

  • Should reconnect replay recent events only, or should the agent notice display absence?
  • Does canvas sleep become a presenceMode, an operator diagnostic, or just a receiver concern?

References

app/api/canvas/delta/route.ts / app/display/display-client.tsx

How Review Should Land

For each question answered: give us your read, a concrete recommendation, and what you would need from us to go deeper. Especially useful: call out where an AGENTV special case is really a missing primitive feature.
14 total questions mirrored from the private review packet.