Agents and Automation: The AI That Works Here, and Its Limits
Last updated: November 20, 2026
Citizen Analyst uses AI in three different ways, and we keep them apart on purpose. This page is the map. It says what each kind of AI worker is, lists every one we describe publicly, and sets out the limits they all share. Each row in the table links to an article that goes deeper.
Our policy is in How Citizen Analyst Uses AI; the rules that bind these workers are in Rules, Gates and Guardrails: How We Constrain Our AI and Ourselves.
Three kinds of AI worker
An agent is a written job description for one AI worker. It names the task, the skills the worker applies and the things it must never do. A human, or a build process a human started, hands it one brief. It works in its own scratch space, writes a results report and stops. It is closest to a staff member given an assignment.
A scheduled job is software that runs on a timer. Nobody starts each run. Some call an AI model and some are plain checks.
A reader-facing AI service runs when a reader asks something, such as the chat on our site. It is the only kind with a public user, so it is the only kind that answers without an editor reading the answer first.
Everything we describe, in one table
| Article | Kind | What triggers it | What it is for | Where a human checks |
|---|---|---|---|---|
| The Oversight Investigation Pipeline Agent | Agent | A human dispatches it, one investigation per run | Builds a sourced chronology, evidence register and draft staff report from public records | A human files every request and edits and promotes the draft |
| The Records Request Agent | Agent | A human dispatches it | Drafts public-records requests and appeals, tracks their deadlines, logs what comes back | A human files every request |
| The Hearing Watch Agent | Agent | A human dispatches it before or after a hearing | Pre-hearing memo and questions; post-hearing transcript analysis and ledgers | A human decides on any witness file or draft |
| The Disclosure Sweep Agent | Agent | A human dispatches it, one subject per run | Gathers public disclosures about one subject and flags patterns for a human look | The reporter decides what is a story |
| The Profile Builder Agents and the Batch Kit | Agents plus a batch kit | A human dispatches one profile, or the kit dispatches builders in waves | Drafts accountability profiles to one standard, alone or in bulk | An editor reviews and promotes |
| The Agency Impact Pipeline Agent | Agent | A human dispatches it, one institution per run | Drafts an Agency Damage Assessment profile | An editor reviews; scoring only on request |
| The Research Worker and the Document Auditor | Agents | Dispatched by an orchestrating session | Gathers sourced research; audits drafts read-only | An editor or builder acts on the findings |
| The Code Builder Agent | Agent (engineering) | Dispatched from a written step brief | Builds one code change in an isolated copy of the repository | The orchestrator verifies; nothing merges on its own |
| News Agent and Threat Tracker | Scheduled job using AI | A timer | Scans news against our marker list and proposes profile updates | Proposed changes wait in a review queue |
| Voting Guide Refresh Agent | Scheduled job using AI | A timer (weekly by default) | Refreshes voting information for 56 jurisdictions | Publishing the refreshed files is the separate, human-triggered publish step |
| The Rights Advisor AI | Reader-facing AI | A reader’s message | Answers civil-rights and voting questions from our articles | Answers are not reviewed one by one; general information, not legal advice |
| Skill Router and Investigation API | Reader-facing AI | A request | Picks the right expert methods for a question; composes investigation plans | It suggests; it does not decide or publish |
| How Publishing and the Chatbot’s Index Stay in Sync | Scheduled jobs | Timers | Keep published articles and the chat’s search index in step; check for drift | Integrity checks alert a human; they do not change content |
| The AI Tools Editors Use in Compose | Editor tools | An editor clicks | Research-to-draft help, scoring, illustrations and more inside our editing console | The editor reads, changes and approves everything |
| Watching Source Pages | Scheduled jobs | Timers | Watch listed source pages for changes | A page is watched only after an editor approves it |
The limits every agent shares
These come from the agents’ written definitions.
- Research goes through one gateway. Agents that research use a single command,
scripts/ca_research.py. It caches results, enforces a per-run budget and keeps a ledger. An agent may not make a network request any other way. A refused request or a spent budget is recorded as a finding, not worked around. - Agents never handle credentials. They do not read the file that holds our keys, never print a key, and never set a switch that turns on a data source that is off by default.
- Agents write only in their own place. The investigation-desk agents write to their own run directory, and at most one draft in
agent-base/knowledgebase/_drafts/. Builder agents edit the draft files they were given. The code builder edits only the files its brief lists, in an isolated copy of the repository. None of them writes into the published knowledge base. - No agent publishes, pushes or merges. Some agents commit drafts, by explicit file path, on their own work branch. None runs a publish command, pushes to a shared branch, merges, or touches the live site. Going live is a separate step a human triggers.
- No agent contacts anyone or files anything. They do not email an agency, committee, witness or company, and they do not submit a request, complaint or tip. A human files every records request. The oversight agent does not start other agents either. It writes recommendations into a handoff note and a human decides.
- Every private person goes through the inclusion gate first. Anyone who is not a public official acting in that capacity, or a company, is checked against
patriot-private-citizen-inclusion-gatebefore an agent names them or builds a file about them. A fail or an unclear result means the person is described by role or left out. See Inclusion and Fairness Gates. - Sourcing rules apply to agents too. An unsourced claim is dropped or tagged for verification. Agents do not invent sources, dates, case numbers or statutes. See Evidence, Sourcing and Language Rules.
- Every run ends with a results file. The agent states what it did and could not do. A missing results file is a failed step.
- An agent stops at the first blocker. If a required skill is missing, a source is unavailable or a gate fails, it reports that and stops that thread. It does not guess past it.
How these limits are held
The limits above are written into each agent’s definition file. A test, tests/test_agent_definitions.py, checks that every definition is well formed, contains no forbidden instruction (such as pushing code or reading the credentials file) and names only script paths that exist. That test is code.
Beyond that, most agent discipline is written procedure that the agent applies and a person checks. Where a check stops a publish in code, the individual articles say so. See AI and Research Runtime Guardrails, Draft Checks and Human Review and Publish-Time Gates.
Agents and the team personas
The Citizen Analyst — Organization Chart shows the specialists we present as the Citizen Analyst team. Those persona pages describe a role and the methods behind it. They are how we present our skills to readers.
An agent is a runnable worker that applies those skills to one task. Each agent article ends with the persona page it mirrors, so you can see the role in newsroom terms. The persona describes the job; the agent is the definition that runs, with its limits written down. Neither is a person, and neither replaces the human editors who decide.
What we leave out on purpose
Our servers are monitored and backed up. We do not describe how, because that detail would help an attacker and not a reader.
How to read the other articles
Each agent article follows one shape, ending with where a human decides and the team persona it mirrors. Skills we publish are linked. Internal working skills are described in words. If an agent does something its article says it may not, tell us; see Corrections & Transparency Policy.
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.
