Related: sources · notes · metadata · Published Pieces

Agent Executes. Radio Abstracts. User Steers. Artifact Records.

The user does not need a diary of the agent’s activity. The user needs an intelligent account of the work.

The mistake is to make the agent narrate itself.

Most agent interfaces are full of process chatter. The agent says what it is about to do, what it just did, what it thinks it found, what command it is running, what file it opened, what test failed, what it plans to try next. This seems useful because it creates a sense of transparency. In practice, it often creates noise.

The user does not need a diary of the agent’s activity. The user needs an intelligent account of the work.

An activity feed says: “I inspected the parser. I found a bug. I edited the file. I ran the test. The test failed. I will now inspect the logs.”

A radio abstraction says: “The parser bug is not local to parsing. It exposes a state-boundary issue. The input is being normalized twice: once before the artifact write and once before the verifier reads it back. The agent is removing the second normalization path and adding a trace check so this cannot silently recur.”

The second is useful because it gives the user the concept.

The basic division should be: agent executes, radio abstracts, user steers, artifact records.

The agent executes because local implementation work belongs to the machine. Read files, search sources, run tests, fetch transcripts, parse PDFs, generate drafts, compare diffs, inspect logs, retry failed commands. This work should not require human attention unless it hits a meaningful boundary.

The radio abstracts because raw process is not the right medium for the user. The radio should surface conceptual bottlenecks, tradeoffs, decision points, failure modes, discoveries, and checkpoints. It should suppress micro-detail until the micro-detail becomes mission-relevant.

The user steers because the human’s role is judgment: taste, priority, mission gradient, tolerance for risk, and recognition of meaning. The user decides when to go deeper, change direction, preserve an invariant, accept a tradeoff, or stop.

The artifact records because memory should not live in the audio. The radio stream is a traversal. The artifact graph is the state. Every important change should land somewhere durable: a vtext, diff, claim graph, event log, source bundle, checkpoint, decision record, or published artifact.

This is the opposite of chat ontology. In chat, the transcript becomes the world. The model’s context window becomes memory. The user scrolls through past turns to reconstruct what happened. Tool calls may be hidden or summarized. Long trajectories become either verbose or opaque.

In Work Radio, the audio is not the state. The audio is the interface to state. If the user interrupts, the system does not lose the thread because the thread is stored in artifacts. The current segment, source path, checkpoint, pending question, and background run all have durable references.

That is why the radio can be calm.

A good Work Radio system should rarely say, “I am doing X now.” It should say, “The important change is X.” It should not narrate every log, read diffs aloud, or produce constant reassurance. It should make uncertainty legible.

The user should hear things like: the agent found a shortcut that would pass the current test but violate the scheduler boundary; the research agent found a stronger opposing frame; the code works in the low-resolution case but the next risk is concurrency; the draft is more readable but less sharp.

These are useful because they expose the work’s shape.

The radio should be editorial, not theatrical. It should not pretend to be a coworker with feelings. It should not fill silence with personality. It should not make the user feel watched or managed. It should provide a clean interface to ongoing transformation.

The artifact does not need flattery. It needs improvement.

The user does not need chatter. They need leverage.

The agent does not need a stage. It needs a landscape.