The State Needs Librarians

Most people meet the state through a form they do not understand, a rule they cannot find, or a decision they cannot appeal. The eligibility rule lives in a PDF that amends a statute from 2009. The denial letter cites a code nobody at the counter can explain. The institution is illegible, and being illegible is its most reliable output.

Public servants face the other side of the same wall. The caseworker and the clerk are not hoarding answers. The knowledge they need is scattered across statutes, case files, three databases, and one person who has been there twenty years and is about to retire.

This is the gap I think AI agents should fill. Not as oracles. As librarians.

The oracle is the wrong dream

The loud version of AI in government is the oracle: ask, get an answer, decision rendered. It is wrong not only because models err, but because an answer you cannot inspect is worse than none when people can be denied housing, benefits, or due process. An oracle that is right ninety-five percent of the time, with no way to see which five percent, is a liability with a confident voice.

A librarian is different, and not passive. A good one interprets the question you actually asked, points you to what matters, and gives a read: start here, this usually governs cases like yours, watch this exception. They have a take. The difference from an oracle is not that they withhold judgment. It is that the judgment is visible and contestable. They show their source, so you can check it and push back.

What a librarian agent does

• Find the right rule. The specific statute or precedent that governs this case, not a plausible paragraph.
• Give a read. An interpretation and a recommendation: a take, not a shrug.
• Trace provenance. Every claim points back to its source, so a human can challenge it instead of trusting a tone.
• Flag what is missing. The form that will bounce, the document not attached, the deadline already passed.
• Surface buried knowledge. The prior ruling, the twenty-year clerk's workaround, found when it is relevant instead of lost when they leave.

The take is the point; so is the limit on it. The agent interprets and recommends. It does not decide the case, file on your behalf, or act on the record alone. In a high-stakes institution, the fastest way to get an agent thrown out is to let it act on its own judgment before anyone trusts it.

The thread in my work

I build AI agents for expert workflows where the output has to be inspected, challenged, and trusted. With Andreas Haupt I co-authored "Bruno: An AI Product Manager for Scientists," a spotlight paper at the ICML AI4Science Workshop; Bruno tracks a team's tasks, decisions, artifacts, and failures without touching the scientific record. Before that I piloted Merton with Stanford teams, and at Sundial Scientific I built interfaces for long-horizon research agents — one turned Sakana AI's AI Scientist-v2 runs into debuggable traces. The goal was never just output. It was to make the work reviewable.

A research team and a benefits office share one disease: knowledge scattered across tools and people, decisions no one can reconstruct, and trust that breaks the moment an answer cannot be traced.

Legible, not automated

The win is not a slicker form or a faster denial. It is a rule you can find, a read you can argue with, and an answer you can trace. The state does not need an oracle that decides for everyone. It needs a librarian: one that shows its work, so the take can be challenged and the judgment stays where it belongs.