The Oversight Investigation Pipeline Agent
Last updated: November 20, 2026
This agent runs one oversight investigation from scope to a draft report, then stops. It works the way a congressional committee’s chief counsel and staff team would, with one large difference: it cannot compel anyone. It lays out what the public record shows, row by row, and leaves every decision to file, contact, dispatch or publish to a human. For how agents fit together, see Agents and Automation: The AI That Works Here, and Its Limits.
What it is
The oversight investigation pipeline (oversight-investigation-pipeline) is an AI agent defined by a written job description. It mirrors a committee investigation’s chief counsel and staff team: scoping the question, planning, preparing records requests, reviewing released material, building a chronology and writing up the findings.
It uses public records only, and uses no committee letterhead or government wording such as “you are directed to” in anything it drafts.
When it runs
A human dispatches it on demand, one investigation per run, when a reporter wants a congressional-style look at an agency, program, company or official. It does not run on a schedule and does not start itself.
What it does
The agent follows a fixed stage order from an internal stage table, described here in our own words.
- Scope. It states the question, lists every majority and minority report already on the matter, and inventories the agency actions those reports cite. It tests each against the “kill criteria”, the findings that would end the inquiry, and marks whether the matter is already settled or only partly corroborated.
- Plan. It completes a plan: the question, competing hypotheses, kill criteria and an owner for each stage.
- Records-request drafts. It drafts each request, choosing federal or state law with FOIA and Sunshine Laws. Every draft is headed “DRAFT: a human files this.”
- Productions and exhibits. It mines publicly released committee productions and hearing exhibits, logging each with a hash, the original page numbers and our control number.
- Chronology. A dated chronology from exhibits and primary records only.
- Evidence register. One row per claim, mapped to supporting exhibits.
- Draft staff report. It writes one draft in
_drafts/, including a section that gives the strongest counter-arguments. - Gates. It runs each integrity gate in a set order and records pass or fail for each.
When a stage finds nothing, the agent writes a “searched, not found” row with the exact query, the source and the date. A search it did not run is labelled “not run”, with the reason. Where the existing record cites hearings, it lists each one and either fetches the transcript through the research gateway or records why it did not.
Evidence rules
- Corroboration. A row needs one primary record, or two independent secondary sources. A weaker item goes into a leads file, never the register.
- Provisional by design. The agent’s source tiers are proposals. The expected result is “pass-provisional” until an editor confirms them.
- Chain of custody. Every document is hashed on intake, keeps its original page numbers beside our control number, and is never altered.
- Labels. Every row says whether it is an allegation, finding, charge, referral or record, and names the body that made it. A referral is not a charge; trade timing is not insider trading; a donation pattern is not a quid pro quo. The agent never implies intent.
- Both sides. Each committee report it cites, majority and minority alike, is tested against its own exhibits.
Budget discipline
The agent sets a research budget first. A rate limit, a refused request or an “unavailable” result is recorded as that source’s outcome and the stage moves on; it never retries in a loop. Court-record and federal-contract lookups are budgeted across everything we run, so it never exceeds its set count. Government audit and inspector-general search pages are link-only: it lists them for a human and does not fetch them.
Skills it applies
Published skills:
- FOIA and Sunshine Laws, for choosing the law and planning the request.
- Congressional Oversight Procedure, to judge whether an inquiry is legitimate and who has jurisdiction.
- Government Oversight Mechanisms, to find which overseer already holds the matter and its open recommendations.
- Citizen Analyst — Private Citizen Inclusion Gate, for every person who is not a public official in that capacity.
- Citizen Analyst — Malice Evaluator (Democratic Malice Assessment), Citizen Analyst Sanity Check and Citation Checker, as gates on the draft.
Internal working skills, described in words only: a staff-report method (stage table, gate order, report structure); records-request drafting; mining of congressional productions and exhibit control; source tiering; quote-fidelity and attribution checks; and two claims-integrity checks on the register.
Rules and gates it obeys
The agent follows Evidence, Sourcing and Language Rules and Inclusion and Fairness Gates. Its runtime limits are in AI and Research Runtime Guardrails. Its draft is subject to Draft Checks and Human Review before anyone can promote it.
Before the plan names anyone, the agent classifies every person in scope as a public official in that capacity, or anyone else. Most witnesses are private employees. A person who fails the inclusion gate, or whose result is unclear, is described by role and never by name in every file. An empty or unresolvable subject is a stop, not a guess.
What it may never do
- File, submit, fax or email a records request, appeal, complaint, tip or referral. A human files every records request.
- Contact a subject, witness, agency, committee office or source. Contact with a whistleblower is a human decision, through secure channels only, and the agent never promises anyone protection.
- Use committee letterhead or government wording, including subpoena language.
- Dispatch, invoke or schedule another agent. Where the The Records Request Agent, The Hearing Watch Agent or The Disclosure Sweep Agent would help, it writes the recommendation and the exact input a human would give, and stops.
- Write into the knowledgebase outside
_drafts/, write more than the one draft, or move, promote or publish that draft. - Read the file that holds our credentials, print a key, or enable a data source.
- Make a network request except through
scripts/ca_research.py. - Commit, merge, push or publish anything.
It stops at the first blocker and says which fired: a required skill that does not resolve, an unresolvable subject, a scope question that cannot be stated falsifiably, a met kill criterion, a missing key for a needed source, or a gate failure it cannot fix inside the draft.
Where a human decides
A human files every records request, dispatches every other agent, edits and promotes the draft, and decides whether anyone is named, any source is contacted or anything is published. Editors also confirm the source tiers behind a “pass-provisional” result. The agent’s results file lists every gate result, inclusion-gate fail, unclean source outcome, recommended handoff and open question. A run with no results file is a failed run.
Team persona
The role is presented on the team page Lucille Thackeray — Oversight Chief Counsel. That page describes the job in newsroom terms; this page describes the runnable agent and its limits.
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.
