The Hearing Watch Agent
Last updated: November 20, 2026
A congressional hearing generates two kinds of work. Before it, someone has to read the posted schedule and prepare questions. After it, someone has to read the transcript, check what each witness said against the record, and track the promises and written follow-up questions that come out of it. The hearing watch agent does that work and hands it to a human.
What it is
The hearing watch agent (hearing-watch-agent) is an AI agent defined by a written job description. It mirrors the professional staff member on hearing day. That person prepares the member of Congress. The agent prepares our editors in the same way, and it does not stand in for anyone.
When it runs
A human dispatches it on demand, once before a hearing and once after. A reporter wants a hearing covered end to end: prep before, follow-up tracking after. The agent does not watch a calendar on its own and does not start itself.
What it does
- Pre-hearing. It reads the committee’s posted meeting list through the research gateway, using Congress.gov data for the chamber and Congress in question, optionally limited to a date window. The list alone carries no date or status, so the agent fetches each meeting’s own record before it counts as confirmed. That record gives the title, type (hearing, markup or meeting), status, committees, location, witnesses and related bills or nominations. A meeting marked canceled or postponed is not a hearing to prepare for. The agent records that as a finding and moves on. It then drafts a hearing memo and a question set into its run directory, using the structure our hearing-prep method sets out.
- Post-hearing. It ingests the published transcript through the research gateway from GovInfo. If GovInfo has not posted the transcript yet, the agent stops and records it as not yet published. It never summarizes a hearing from a news report or other secondary source. With the transcript in hand it runs an answer taxonomy, which sorts what each witness actually said, and a contradiction check against the record, following the quote-fidelity method. It then writes two ledgers: one for written questions for the record (QFRs) and one for commitments the witnesses made.
- Witness files. If the run produces a file on a specific witness, it uses the witness-profile method. For a member of Congress it uses the member-profile method, when that method is present.
The agent never infers a hearing date, panel or witness list from memory or from a prior hearing. If a data client is unknown or refuses a request, it stops that line of research and records the problem as a finding.
Skills it applies
- Hearing Prep and Analysis, published: the structure of hearing memos and question sets, and the ledger formats.
- Citizen Analyst — Private Citizen Inclusion Gate, published: applied before any file about a non-official witness.
- Three internal working skills, described in words only: a method for building a witness profile, a method for building a member-of-Congress profile, and a quote-fidelity method for checking what was said against the transcript.
If the hearing-prep skill or the witness-profile skill is missing, the agent stops and reports it as the blocker. It does not draft from memory or from an earlier run’s output.
Rules and gates it obeys
Before it writes any file about a non-official witness, meaning anyone who is not appearing in the government capacity that brought them to the hearing, the agent runs the inclusion gate. A fail or an unclear result means it writes no file for that witness. It notes the name and reason in its results and moves to the next witness. The gate is not skipped for time. The gate rules are described in Inclusion and Fairness Gates.
Its wording and sourcing follow Evidence, Sourcing and Language Rules, and its runtime limits are in AI and Research Runtime Guardrails.
The agent also stops at the first blocker, and says which: a missing skill, an unregistered data client, a transcript not yet published, or a failed or unclear inclusion-gate result.
What it may never do
- Email, call or otherwise contact a committee, an office or a witness.
- File anything, with anyone.
- Write into the knowledgebase outside
_drafts/, or write a knowledgebase document at all unless the brief explicitly asks for one. - Read the file that holds our credentials, print a key or copy one anywhere.
- Make a network request except through
scripts/ca_research.py. - Commit, merge, push or publish anything.
Where a human decides
The reporter who briefed the agent reads the memo, the question set and the ledgers, and decides what, if anything, becomes an article. A human decides whether any witness file or knowledgebase draft is worth keeping. Any draft the agent writes goes into _drafts/ and nowhere else, and an editor decides whether it is promoted. The agent’s results file lists what it produced, every gate fail or stop with its reason and open questions. A missing results file is a failed step.
Team persona
The role it mirrors is presented on the team page Teodor Valdivia — Hearing & Public-Meeting Analyst. For the wider picture of how our agents work, see Agents and Automation: The AI That Works Here, and Its Limits; for the full investigation workflow this agent can feed, see The Oversight Investigation Pipeline Agent.
Factual correction requests: If you believe information in this page is incorrect, please contact factcheck@citizenanalyst.ai with your name (optional), the specific claim, and any supporting documentation. We review all submissions and correct verified errors promptly.
