SeedRealtime Full-Duplex Phone Agent Architecture: Voice Is Only Half the Stack

What SeedRealtime and Seeduplex Change for Phone Agents

SeedRealtime and Seeduplex are important because they move voice AI closer to natural conversation. In ByteDance Seed's April 9, 2026 introduction to Seed Full-Duplex Speech LLM, Seed describes Seeduplex as a native listen-while-speaking framework for more natural speech interaction, with improvements in attentive listening, interference suppression, endpointing, and interruption handling. Seed also describes rollout in Doubao.

The SeedRealtime model page places SeedRealtime in the current Seed model family, which gives builders a clearer entity label for this line of real-time voice work. For phone-agent builders, the meaningful shift is not only faster speech. It is the move from turn-by-turn voice commands to an interaction model where the user can interrupt, correct, hesitate, or speak over the assistant more naturally.

From our FoneClaw builder perspective, that interaction breakthrough is exciting because phones are messy, mobile, and interrupt-driven. A user may start a request while walking, change their mind mid-sentence, speak over the assistant, or correct a contact name before the next step. Full-duplex voice AI can make that front end feel more human. The phone still needs a governed Android execution layer after the model understands intent.

For the wider interaction thesis behind voice-first phones, Voice-First AI Phone Interaction: Why the Next Phone Starts With Intent explains why future phones should begin with the user's goal, not the app grid.

How Full-Duplex Voice AI Handles Pauses, Noise, and Interruptions

Most voice systems people know are half-duplex in practice. The user speaks, the system waits for an endpoint, the model generates, then the assistant speaks back. That rhythm works for simple commands, but it feels rigid when the user pauses, adds a correction, speaks over the answer, or reacts to something happening on the phone.

Full-duplex voice AI changes that rhythm by listening while speaking. The assistant can keep generating audio while also monitoring incoming speech. That makes barge-in possible: the user interrupts with "wait," "use Alex at work," or "make that 3:30," and the system can route the new audio into the conversation instead of forcing the user to wait for the assistant to finish. Seed attributes its Seeduplex gains to a native framework designed around attentive listening, robust interference suppression, better endpointing, and interruption handling.

Noise and side speech are the hard parts. A phone agent operates around traffic, kitchens, open offices, family rooms, Bluetooth audio, and app sounds. The model has to decide whether incoming audio is a correction, a new command, background speech, or acoustic interference. The paper How Should LLMs Listen While Speaking? explains the routing problem directly: full-duplex systems must handle user audio that arrives during generation, and channel fusion or cross-attention routing creates tradeoffs between grounding and context robustness.

Latency matters too. A model can be smart and still feel awkward if it responds late after an interruption. For readers following the model-speed side of phone agents, 1000 TPS LLMs and the Phone Agent Era: What Speed Really Changes covers why raw generation speed helps only when the rest of the task stack can use it.

Why Conversation and Android Execution Are Separate Planes

A full-duplex model improves the interaction plane. It helps the assistant hear, speak, pause, recover from interruptions, and stay conversational while the user changes direction. Android execution lives on another plane. It handles apps, permissions, state, tools, approvals, device settings, contacts, files, messages, maps, and visible results.

Collapsing those planes creates accidental-action risk. A user might interrupt with "send it," then immediately add "wait, not that Alex." The conversation model has to capture the correction, but the execution layer has to know that sending a message is an external effect. It needs a recipient, complete body, default app, stable Send control, and review step. Natural speech alone is not the approval record for a sensitive phone action.

This separation is central to how we build FoneClaw. The model can understand the user's intent and update the plan in real time. Governed tools carry out supported Android actions only through explicit task contracts. The execution layer keeps task state, permission needs, approval status, observed results, and recovery visible. That is what makes a real time voice agent useful on a phone instead of merely impressive in a demo.

For the general execution model, AI Agent Phone Control: How Android Phone Agents Turn Intent Into Action explains how Android phone agents connect intent to supported device actions with visible control.

A Seven-Stage Contract From Spoken Intent to Verified Result

The practical Android phone agent architecture needs a contract between conversation and action. We use seven stages when thinking about this problem: capture, clarify, plan, approve, execute, verify, and recover. Full-duplex voice AI can improve the first two stages, but the later stages make the phone trustworthy.

1. Capture: receive the user's spoken request and keep listening for corrections, interruptions, and changes of intent.
2. Clarify: resolve uncertain names, apps, time ranges, accounts, contacts, or screen references before action begins.
3. Plan: choose a supported route and separate conversation work from device execution.
4. Approve: show the target, content, permission, and consequence when a phone action affects another person, account, file, setting, or private context.
5. Execute: call governed Android tools or service capabilities only for the approved task.
6. Verify: check the observed result, such as a setting state, visible draft, opened app, navigation handoff, or completed action.
7. Recover: keep the task understandable when permission is missing, the screen changes, the app blocks a route, or the user interrupts.

State and approval have to persist across those stages. If the user corrects the instruction during dictation, the pending approval should update. If a permission prompt appears, the task should remain linked to the original intent. If the model hears a new command while speaking, the execution layer should decide whether that new input changes the existing task, pauses it, or starts another one.

That is why approval design deserves its own product treatment. AI Agent Approval UX on Phones: Confidence, Rationale, and Recovery goes deeper on how a phone should show the reason, target, consequence, and recovery path before a meaningful action.

How We Are Building the Current FoneClaw Execution Layer

Today, SeedRealtime and Seeduplex are architectural context for this article. FoneClaw uses our current configured-model and governed-tool stack; there is no official SeedRealtime-FoneClaw integration today. That distinction matters because the useful lesson from full-duplex research is architectural: fluid conversation needs a reliable execution layer before it becomes dependable phone work.

The current FoneClaw release information gives us the current execution baseline. It adds floating-assistant access, same-phone task continuity, permission recovery, and quick actions. Voice input is user-triggered. Current-screen context is also user-triggered, and the floating assistant can attach the current screen on demand while excluding FoneClaw overlays. That gives the user control over when the agent receives visible context.

We build FoneClaw around governed Android actions. Supported workflows include selected Do Not Disturb changes, visible messaging flows, dialer preparation, navigation handoff, and selected Android settings actions. Sensitive external effects stay visible and reviewable. Missing permissions become recovery steps. Unsupported screen states become visible takeover points. The current product capability areas are summarized on FoneClaw Features.

The floating assistant is especially relevant to the full-duplex conversation future. A user can remain inside another app, call FoneClaw, attach context, ask for a task, approve or stop, and recover through the same phone task. Android Floating AI Assistant: Use Current-Screen Context Safely explains that current-screen path in detail.

Full-Duplex Phone-Agent Scenarios That Need Execution Safeguards

Consider meeting mode. The user says, "Turn on Do Not Disturb until this meeting ends," then interrupts the assistant with, "actually allow calls from family." A full-duplex model can catch the correction while speaking. The execution layer still has to prepare the supported setting change, show the exception, ask for approval where needed, apply the change, and verify the resulting state.

Messaging is another useful example. The user dictates, "Text Morgan that I will arrive at 4," then says over the assistant, "wait, make that 4:30 and use Morgan from the project team." The interaction layer handles interruption and correction. The execution layer checks the recipient, body, default messaging app, and stable Send control before a plain-text SMS leaves the phone. If the screen is ambiguous, the task remains visible for user completion.

Calls and navigation show the value of handoff. A user says, "Call the clinic, then navigate there," and corrects the destination while the assistant is speaking. The model should keep the conversation fluid. The phone agent should prepare the call path, hand navigation to the selected map app, and make clear which step is waiting, running, or complete.

Interrupted multi-step work is where a real time voice agent becomes a product system. If permission is missing, the user denies approval, the app changes screen, or Bluetooth audio interferes, the task should stay recoverable. That is the difference between a natural demo and a reliable Android phone agent architecture.

Privacy, Battery, Audio-Visual Limits, and What to Build Next

Full-duplex voice AI brings new constraints. Listening while speaking can improve natural interaction, but microphone behavior must be legible to the user. Battery and latency matter because always-hot audio processing can become expensive on a phone. Noise suppression has to distinguish the user's correction from background speech. Interruption handling has to prevent accidental actions when the user speaks casually near the device.

Visual grounding is a separate frontier. Seed's April 2026 release describes visual input and proactive interaction as future work. The VideoFDB full-duplex audio-visual benchmark also shows why streaming audio-visual grounding remains its own evaluation challenge, with tested systems often underusing visual streams outside explicit visual questions. Listen-see-speak interaction needs more than a voice model label.

For FoneClaw, the builder checklist is practical. First, keep voice and screen context user-triggered. Second, preserve task state across interruption. Third, separate model conversation from governed execution. Fourth, require visible approval for meaningful external effects. Fifth, verify the result instead of assuming completion. Sixth, offer recovery when permissions, app state, noise, or context fail. Seventh, keep battery and privacy costs visible in product design.

That is where the next generation of phone agents should go. Full-duplex models can make the conversation feel alive. Governed Android execution makes the result dependable. FoneClaw's path is to keep improving the execution plane so future real-time models can become useful phone agents, not just better talkers.