How the agent factory works
You give Campaign Factory one UK local or public-policy problem and a named place. A team of AI agents researches it live and assembles a whole campaign in front of you. This page explains the machine behind that — in plain terms, and honestly about what it can't do.
The factory is a team of AI agents that builds a campaign the way a good campaigns office would: research first, then evidence-checking, then analysis, then strategy, then the materials. Each agent has one job and a named contract — a research director, an evidence adjudicator, a power-map analyst, a strategy architect, three producers, and so on. Thirteen roles are fixed; on top of those, the factory picks specialists to suit your problem — council records for a council decision, planning for a planning fight, parliamentary for anything that runs through Westminster.
It is a factory in the literal sense: work passes along a line, one stage at a time, and a reviewer stands between the stages. Nothing reaches you that hasn't been through the gates.
The first half: finding out
Research starts wide: the research director scopes the problem and sets the questions, while the specialists dig into the local record at the same time. Evidence then narrows: the adjudicator takes every claim the researchers surfaced and decides what it can honestly be called — confirmed, qualified, conflicted, not found, or stale.
The middle: making sense of it
Analysis runs four agents side by side: one sets the objective, one maps the formal decision route, one builds the power and stakeholder map, one works out where pressure could actually bite.
The second half: deciding and making
Strategy is a single architect that turns the analysis into one coherent approach — and if the reviewer isn't satisfied, it sends the strategy back for one round of revision before anything gets built on it. Production then runs five agents at once: tactics, organising, and the three producers who draft the lobbying, media and digital materials.
The gates
Between every stage sits the same reviewer, reading what was proposed before the next stage is allowed to build on it. More on that in section 4.
The researchers work from the real public record — council papers, official announcements, local reporting — through a bounded budget of live web searches. There is no synthetic data anywhere in the product: the original prototype had a fake-campaign generator, and it was deliberately not carried over. Every claim that survives research is put in front of the evidence adjudicator, which assigns it exactly one of seven labels:
The labels follow the claims all the way into the drafts: an unresolved fact in a letter or press release appears as a highlighted [VERIFY: …] placeholder rather than a confident-sounding guess, so nothing unchecked can slip into an email unnoticed. And if an agent ever returns a label that isn't on the list, it is downgraded to “Verification incomplete” — never upgraded.
The reviewer gate
A dedicated reviewer reads each stage's output and does one of three things: accepts it, returns it once for revision, or rejects it. It writes a short report on every stage it closes, and when it disagrees with something, the disagreement is preserved in the record rather than smoothed over. A rejected piece of work stays rejected — it is shown as such, not quietly patched.
Decision points — surfaced to you
Some choices shouldn't be made by a machine on your behalf. When the factory hits one — an ambiguous scope, genuinely conflicting evidence, a real strategic fork, or something only local knowledge can settle — it raises a judgement request: a plain question, the options, and the default it will proceed with.
You can answer, accept the default, or leave it — the line keeps moving either way, and whatever you choose is applied to everything it affects. The point is that the choice is visible: the factory tells you what it assumed instead of hiding it.
And after the run
Everything the factory produces is a first draft addressed to a person. Nothing is sent, published or acted on unless a human has edited and approved it.
Campaign Brief
Objective and Theory of Change
Power and Stakeholder Map
Campaign Strategy
Tactics and Timeline
Organising Plan
Lobbying Pack
Media Pack
Digital Campaign Pack
Each document finishes as ready or needs verification — an honest status, not a decoration. The packs carry their unresolved facts as visible [VERIFY: …] notes.
Every campaign runs under a hard time limit and cost guards, checked before each stage is allowed to start. When a limit trips, the run finishes deterministically: what's done is kept, and whatever didn't happen is recorded as a terminal gap— a visible entry saying “this work was not completed”.
A run that stops early is shown as Stopped early, with everything it finished kept. A run that runs out of time is graded on what it actually built — Nearly complete, or a plain count like 6 of 9 sections built — never rounded up to Complete. The factory never fabricates completion — that rule is enforced in the machinery, not just promised in the copy.
What it can't do
- It can only verify what's on the public record — if the council hasn't published it, the factory can't confirm it.
- Positions attributed to named people are inferences to be checked with them, never facts.
- Time-capped research can leave questions open; they are listed as unresolved, not answered by guesswork.
- It doesn't know your community. Local knowledge, relationships and political judgement stay human.
The repository holds two versions of the same idea, sharing the same evidence rules. The first is a straightforward pipeline — a handful of large model calls in sequence — which shipped first and still runs as the legacy builder. The second is the multi-agent factory described on this page: a genuine fifteen-agent graph (built on the open-source LangGraph) running on its own always-on worker, saving its progress at every stage and streaming its work to your browser as it happens.
The factory rewrite was itself built by AI coding agents working to human decisions — an AI agent team building an AI agent factory. Each significant choice was argued out and written down as an architecture decision record before it was implemented, so the reasoning is as inspectable as the code.