Remote Frog

We talk a lot about whether an AI agent can answer a question, but not enough about whether the material it reads is actually readable for the job we expect it to do.

That is where reading usability for agents comes in.

My short version: documentation, schemas, and knowledge bases are only good if they help an agent to answer questions correctly.

This matters because agents do not magically solve ambiguity. They operationalize it. If a definition is fuzzy, if a metric changed three months ago and the explanation now lives in a Slack thread, if a table looks documented but hides its real meaning in pipeline code, the agent will often still produce something that looks plausible. Dangerously plausible.

So when I say reading usability, I do not mean “can the model parse the text?” I mean something closer to this:

• Can the agent identify what source is authoritative?
• Can it tell which terms are defined and which ones are not?
• Can it distinguish monitoring material from explanatory material?
• Can it verify an answer instead of merely assembling a very elegant guess?

And this is where the dashboard analogy is useful. In Storytelling with Data there is a very healthy distinction between dashboards for exploration and communications for explanation. A dashboard helps you monitor, explore, and notice patterns. A presentation or focused narrative helps you communicate a specific insight and drive action. If you confuse the two, you make life harder for humans.

Exactly the same thing happens with agents.

Some artifacts are built for exploration: broad documentation hubs, warehouse schemas, internal wikis, dashboards, collections of links, and long project docs. These are useful, but they are not automatically usable by an agent asked to provide a verifiable answer. Other artefacts are better suited to explanation: a carefully scoped FAQ, a source-of-truth spec, a decision record, a metric definition with caveats, a tested answer pipeline.

So the question is not only “does the agent have access to the content?” but also “what kind of reading task am I asking it to perform?” If I want a trustworthy answer, I should not be surprised that dumping an exploratory landscape onto an agent produces exploratory-quality certainty.

For me, reading usability for agents has at least four ingredients:

• Legibility: terms, relationships, caveats, and scope are explicit.
• Authority: the agent can tell what source should win when sources disagree.
• Verifiability: the path from answer to evidence is inspectable.
• Task fit: the artifact matches the kind of reasoning we expect, whether exploration, monitoring, explanation, or decision support.

Humans compensate for poor reading usability with context, memory, habit, and complaints. Complaints are underrated. Agents do not complain. They just keep going.

That is why I think this topic matters. If we want verifiable answers from agents, we need to care not only about models, prompts, and tools, but also about whether what they read is designed to be understood, checked, and used for the kind of answer we are asking for.

In short: reading usability for agents is the equivalent of usability in dashboards. If people cannot find the question, the signal, or the next step, the dashboard fails. If an agent cannot find the authority, the meaning, or the evidence, the answering system fails. Same song, different dance floor.