Related: sources · notes · metadata · Published Pieces

Audio Runway Buys Cognition Time

The graph gives the radio runway. The runway gives agents time. The agents update the graph.

The central advantage of automatic radio is temporal.

In chat, latency is dead time. The user asks a hard question, then waits. If the model thinks for three minutes, the product feels broken. If a research agent needs thirty minutes, the user leaves. If a coding agent needs four hours, the user cannot sensibly remain in the thread.

Audio changes the economics of waiting.

If the system has an artifact graph, citation network, prior vtexts, source clips, summaries, human recordings, claim maps, and cached transforms, then a single prompt can produce immediate runway. The system can begin with what is already known while deeper agents work in the background.

That runway buys cognition time.

A user asks to catch up on the voice AI landscape and what it means for Choir Radio. The system does not need to wait until every research path finishes. It can start with orientation: realtime audio models, cascaded STT-text-TTS pipelines, and whether automatic radio is better understood as a voice agent or as media traversal over an artifact graph.

Then it can traverse existing material: the current position that voice should be thin, intelligence should remain in text-native multiagent systems and artifact graphs, and realtime presence should not own memory.

Now background agents have time. One searches open-source realtime voice models. Another compares STT and TTS components. Another reviews competitors. Another drafts an implementation plan. Another extracts the cost model. The radio keeps going.

A few minutes later, the research agent updates the stream: public evidence still favors hybrid pipelines for reasoning-heavy tasks. Native speech systems are improving, but they often trade depth for immediacy. This supports the current Choir architecture.

The user never stared at a spinner.

This is the product insight. Automatic radio can make long inference feel alive because the user is already receiving value while the system computes. The listening stream becomes a buffer between human attention and machine time.

This only works if there is reusable cognition. A shallow voice agent cannot do it because every answer has to be generated from the current conversation. A strong artifact system can do it because the graph already contains paths: prior work, related claims, source trails, objections, human clips, cached summaries, known distinctions.

The richer the graph, the longer the runway.

A mature Choir system should be able to speak for thirty minutes on a serious topic using existing structure while new agents deepen the answer. That does not mean rambling. It means moving through layers: orientation, context, sources, prior art, disagreement, synthesis, unresolved questions, and current work.

The user can interrupt at any point: skip the background, go deeper on the benchmark, play the original clip, what did I say about this last week, how does this affect the build. The interruption becomes an event. The radio branches, answers, and returns.

This makes audio different from text. Long text output often burdens the user. Long audio, if well produced and interruptible, can be the ideal medium. People already listen for hours to podcasts, interviews, lectures, and radio. The problem is not duration. The problem is signal.

Audio runway works when the stream remains source-grounded, paced, and responsive. It fails when it becomes filler. The system must always know what it is doing: orienting, explaining, contrasting, quoting, updating, or asking for a decision.

The background agents should not be hidden in the sense of unaccountable. Their results should land in artifacts. The radio may summarize, but the artifact records. The user can later inspect the sources, vtext, code diff, checkpoint, or decision record.

That is the full loop: the graph gives the radio runway; the runway gives agents time; the agents update the graph; the radio reports what changed; the user steers; the artifact remembers.

This is why automatic radio is not merely an audio interface. It is the temporal layer of an automatic computer.