Transmission // 21
Legibility is a feature
1 min
An agent that cannot show its work is not autonomous, it is unaccountable. The difference is entirely in the tooling around it.
The pitch for agents is that you stop watching. The reality is that you stop watching the steps you already trust, which is a much smaller claim and a much better one.
Getting there requires that the steps be watchable in the first place. Most harnesses treat the trace as debug output: verbose, unstructured, discarded on exit. That is backwards. The trace is the product. The answer is just the last line of it.
What “legible” actually means
It is not a log level. A legible run answers four questions without you reconstructing anything:
- What went in. Not “the prompt” — the whole context: which files, which tool definitions, which prior turns survived compaction.
- What it chose. Every tool call, with arguments, in order, including the ones that failed and the retries that followed.
- What it cost. Tokens per step, cache hits, wall time. Aggregates hide the one step that burned half the budget.
- What changed. The diff on disk, separated from the narration about the diff.
Miss any one and you are back to trusting a summary written by the thing you were trying to audit.
The uncomfortable part
Legibility is expensive to build and boring to demo. Nobody screenshots a good trace viewer. But it is the only thing that converts “the agent did something strange” from a shrug into a bug report, and a bug report is the only artefact that makes the next run better.
You cannot steer what you cannot see. Everything else is vibes with a spinner.