Bandwidth Inversion
The structural condition in which a human and an agent have opposite communication-efficiency profiles: the human receives fastest through vision and emits fastest through speech, while the agent receives and emits fastest through structured text and treats higher-dimensional channels such as audio and video as progressively lossier — so that no single channel is optimal for both parties to a human-agent exchange.
A human and an agent do not merely differ in speed. They differ in which channels are fast for them, and the difference runs in opposite directions. For a human, vision is the highest-bandwidth input — a screen conveys state faster than any description of it can be read or heard — and speech is the highest-bandwidth output, several times faster than typing. For an agent, structured text is both: it is parsed natively, carries no ambiguity of rendering, and can be read and written at the rate of the underlying model. Audio costs the agent a transcription step. Video costs it far more — the reconstruction of meaning from a stream that is high in dimension and low in structure. Bandwidth Inversion is the name for that opposition. Where the human's ranking runs vision, speech, text, the agent's runs text, speech, video, and the two curves cross rather than align.
The condition has a direct consequence for the corpus's existing claims about the Steward. The Steward Cannot Govern What They Cannot See and The UI That Runs the Business established that the Steward's interaction surface is load-bearing, and The Steward's Blind Spot showed the metric consequence of getting it wrong: an unusable surface produces disengagement, and disengagement produces Nominal MTTI. None of those memos stated why a given surface is usable or not. Bandwidth Inversion supplies the reason. A surface that exposes the agent's native medium to the Steward — raw structured state, logs, a ledger — hands the human their slowest input channel. A surface that asks the Steward to emit in the agent's native medium — typed structured commands — hands the human their slowest output channel. Both produce the blindness and friction the earlier memos measured, for a mechanical reason rather than a design failure.
The complication is that the inversion is not symmetric in cost. The human cannot be redesigned; the agent can be given a translation step. That asymmetry is what makes Bandwidth Inversion a design condition rather than a limit — it locates where the translation work has to sit, on the system side, and names the discipline that does it: the Transcoding Layer. The Machine-Readable Interface already solves the mirror problem, making a business legible to agents. Bandwidth Inversion states why the reverse direction, making an agentic system legible to a human, cannot be solved by the same move.
Arco treats the specific channel rankings as a current reading. Models get faster at audio and video; humans do not change. The direction of the asymmetry may narrow on the agent's side. The inversion itself does not close, because one of the two curves is fixed.
Application
Before specifying any Steward-facing surface, score each channel twice — once for the human's receive and emit speed, once for the agent's — and design against the two rankings rather than picking the channel that is convenient for one side.
Context
The principle underneath Steward Experience and the Audit Surface: it explains why an interface built on the agent's native medium leaves the Steward blind, and why an interface built on the human's native medium is one the agent cannot read — the missing mechanism beneath the corpus's claim that the Steward's surface is load-bearing.
This term is machine-readable
Any MCP-compatible AI assistant can retrieve the canonical definition of Bandwidth Inversion at inference time — no training approximation.
Related Terms
First used: September 2026
Edition 1 · updated September 2026