The Agency Impact Pipeline Agent
Last updated: November 20, 2026
An Agency Damage Assessment documents the harm done to one federal institution since 2025, and sets out how each harm could be reversed. It is a different instrument from a person profile, with its own method. The agency impact agent builds the draft for it, researches the record, and stops short of anything that needs an editor’s judgment.
What it is
The agency impact pipeline (agency-impact-pipeline) is an AI agent defined by a written job description. It builds or advances an Agency Damage Assessment profile for one federal institution: an agency, department, bureau, office or chartered body that the Trump administration or DOGE has harmed. It mirrors the agency damage assessor on the team. The method is the published skill Agency Damage Assessment.
When it runs
A human dispatches it on demand. A request to create, research, stub, score or update an agency impact profile, to build an agency’s personnel ledger, or to draft reversal pointers will start it. It does not run on a schedule.
What it does
- Gate. It confirms that the subject is a harmed institution and not an actor body such as a task force, a DOGE team, a fund or a tool. If it is an actor body, the agent stops and says which agency profile it belongs in, as an instrument of harm.
- Scaffold. It chooses a plain kebab-case slug, checks that it is unused with a draft-aware lookup and writes a stub in
agent-base/knowledgebase/_drafts/. The stub isstatus: draft, in the organizations category, marked as a stub. An existing profile is edited in place, never copied into a dated sibling. - Research. It reads committed primary-record data first, before any search: budget figures from USAspending, workforce data from the Office of Personnel Management and apportionment data from the Office of Management and Budget. Those files are evidence, never a score. Then it reads the evidence store, where records already fetched for the agency live, and cites those before fetching again. Every network request goes through the research gateway,
scripts/ca_research.py, under a budget. A refused request is a finding. - Personnel Ledger. It records who was fired, pushed out, placed on leave or resigned. It names only senior or publicly reported people and counts rank-and-file losses in aggregate. Each row carries the action type, date, source link and the capability lost. It links an existing person profile where one exists.
- Assessment. In a stub, it marks each dimension “unscored”, adds one or two evidence bullets and lists the evidence still needed. It scores only when a human asks for a scored profile, and only at the evidence tier the methodology requires.
- Reversal pointers. For each dimension: the class of remedy, what should be preserved now, and the time horizon.
- Check. It confirms that the frontmatter parses, the category exists, every wiki link resolves and there are no Obsidian-only constructs. It runs the draft check (
scripts/check_draft.py) and the case-caption check. - Report. It lists the open blockers. It commits only if asked, and only in its own isolated copy of the repository, by explicit path.
Facts are handled strictly. Every claim is cited. A claim from a single source reads “According to…” Each web result is tied to this agency and to the 2025-to-present period before it is used. The agent never states a person’s motive. It records the action and its documented effect.
Skills it applies
- Agency Damage Assessment, published: the eight scored dimensions, the personnel ledger and the reversal pointers.
- A methodology document that governs the gate, the evidence tiers, the anchors, the ledger, the reversal pointers and the claims-integrity gate. The overall approach is summarized in Scoring & Inclusion Systems — Complete Inventory.
- An agency impact profile template.
Rules and gates it obeys
The agency instrument is one of several, and each measures a different unit. This one measures one federal institution, and it uses the findings of other instruments without re-scoring them. See Scoring & Inclusion Systems — Complete Inventory for how they differ. The wording and sourcing rules in Evidence, Sourcing and Language Rules apply, and the runtime limits are in AI and Research Runtime Guardrails.
The agent stops at the first unresolved blocker rather than guessing past it. A stub is never scored, which is the clearest gate in its job.
What it may never do
- Score a stub.
- State a person’s motive, or name rank-and-file employees. Senior and publicly reported people only; everyone else in aggregate.
- Promote, publish, push, merge, open a pull request or run the GitHub tool.
- Commit unless asked, and then only in its own copy of the repository, on its own branch, by explicit path.
- Make a network request except through
scripts/ca_research.py, or read the file that holds our credentials.
Where a human decides
A human decides whether a profile is scored at all, and an editor reviews it afterwards. Shipping a merge is the orchestrator’s step, never the agent’s. An agency that fails the gate goes back to the human with the reason. The agent’s open-blockers report is the handoff.
Team persona
The role it mirrors is presented on the team page Marisol Hartwell — Agency Damage Assessor. See The Profile Builder Agents and the Batch Kit for how person profiles are built, The Research Worker and the Document Auditor for independent checking, and Agents and Automation: The AI That Works Here, and Its Limits for the whole set.
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.
