By Mark Preston
Update — April 2026. The debate tool above has evolved. What began as three agents arguing a motion is now a full Oxford Union bench: four speakers a side, each with a distinct role and a distinct historical voice. The post below has been rewritten to describe where the architecture has actually ended up — and why it matters.
Imagine, for a moment, that you could fill an Oxford Union bench with anyone you wanted.
On the proposition side: Christopher Hitchens opens — rolling sentences, literary bombardment, moral certainty delivered with theatrical élan. Edmund Burke builds — organic metaphors, accumulated wisdom, arguments that unroll slowly and arrive with the force of inevitability. Malcolm X takes the clash — controlled fury, refusal of the opposition’s framing, rhetorical questions that have no comfortable answer. Winston Churchill closes — architectural sentences that ascend to a moral crescendo.
On the opposition side: Margaret Thatcher, William F. Buckley Jr., Emmeline Pankhurst, Cicero.
The motion: This House Believes That Artificial Intelligence Will Be Humanity’s Last Great Invention.
You can now watch that debate run in a browser window. Not as a gimmick — as a serious piece of AI architecture pretending to be a party trick. This post is about the architecture.
What Is a Mixture of Agents?
Most people’s experience of AI is a single conversation: you ask, it answers. One model, one response, one perspective.
Mixture of Agents (MoA) is different. Instead of asking one AI for an answer, you ask several — each reasoning independently, each approaching the problem from a different angle. Their outputs are then synthesised into a combined response that is, demonstrably, better than any single agent could produce alone.
Think of it less like asking a consultant and more like convening a committee — except this committee doesn’t suffer from groupthink, doesn’t defer to seniority, and can run in seconds.
The key insight: diversity of reasoning produces better answers than depth of reasoning alone. An agent with a slightly different framing of the problem will catch things the first agent missed. An adversarial agent — one specifically tasked with finding holes in the argument — will surface weaknesses before they become failures.
Good organisations already know this. You don’t present a plan and ask if everyone agrees. You present it and ask someone to break it. Formula 1 teams don’t debrief a race with one engineer. They debrief with a strategist, a vehicle dynamicist, a tyre engineer, and a race engineer — each looking at the same stint through a different lens, and the sum of those lenses is sharper than any one of them on their own.
From Three Voices to a Full Bench
The first version of this tool — the one that went live on this page in February — used three agents: Proposition, Opposition, and a Devil’s Advocate that attacked whichever side was winning. That already produced debates noticeably richer than a single-model answer. But it had a structural lie baked into it: the Oxford Union isn’t two voices. It’s a bench.
At the Union, each side fields three or four speakers, each with a distinct job.
- The Opener defines the terms, sets the burden of proof, and plants the flag.
- The Builder deepens the case with evidence, answers the opposition’s opening, and extends into a second substantive theme.
- The Clash engages the central disagreement head-on, repairs damage, and holds the line.
- The Closer crystallises the debate into voting issues, weighs the two cases, and delivers the peroration.
The new version — branded DeQuorum and sitting live above — models that bench properly. Each speaker is its own agent with its own role constraints, its own prompt structure, and its own place in the sequence. The Opener is not allowed to trespass on the Closer’s job. The Closer is not allowed to introduce new substantive material. The Clash speaker is forbidden from avoiding the main disagreement. These aren’t stylistic preferences — they’re hard rules in the prompt, because a real Union debate is a structured game.
And the speeches are generated in sequence, each speaker seeing what came before, each one knowing what their role expects of them and what the next speaker is waiting to receive. Speeches stream live into the bench roster as they generate — you can watch the debate unfold in real time rather than wait for a monolithic paragraph to drop at the end.
The Persona Library: Thirteen Great Debaters
The architectural move that makes the bench interesting isn’t the roles. It’s the personas.
Each speaker on each side can be assigned an identity — a famous debater whose rhetorical DNA is injected into the agent. Not as a costume, but as an instruction. Hitchens does not argue like Obama. Thatcher does not concede like Beauvoir. A persona encodes not just what an agent says but how it reasons.
The current library has thirteen:
- Christopher Hitchens — polemical wit, literary bombardment
- William F. Buckley Jr. — conservative forensics, aristocratic diction
- Malcolm X — moral urgency, refusal of the opposition’s framing
- Margaret Thatcher — iron logic, relentless enumeration
- Barack Obama — professorial bridge-building, concession-and-pivot
- Emmeline Pankhurst — suffragette moral force, personal testimony as evidence
- Edmund Burke — philosophical conservatism, organic metaphors
- Abraham Lincoln — biblical cadence, stripped-down moral clarity
- Cicero — full classical rhetoric, devastating peroration
- Frederick Douglass — abolitionist pathos fused with constitutional rigour
- Mary Wollstonecraft — Enlightenment rationalism, demolition by reason
- Winston Churchill — soaring oratory, architectural sentences
- Simone de Beauvoir — existentialist analysis, refusal of false dichotomies
Each persona is defined with an era, a nationality, a catalogue of notable debates, a style profile (rhetorical devices, temperature hint, verbosity), and — crucially — a prompt-injection block that is pasted verbatim into the agent’s instructions. Each persona also has a suggested role or two where their style plays strongest. Burke wants to build. Churchill wants to close. Hitchens wants to open or handle clash. Beauvoir is at her best late in the debate, crystallising the philosophical scaffolding.
The library is deliberately cosmopolitan — British, American, Roman, French, Irish; 70 BC to present; conservative, liberal, radical, philosophical. The point isn’t who you agree with. The point is that different rhetorical traditions surface different arguments. A Hitchens opener on an AGI motion finds different weaknesses in the opposition than a Thatcher opener will. Stack the wrong bench for the wrong motion and the debate is thinner. Stack it right and you get an argument that no single model — and arguably no single human — could produce alone.
Why This Matters Beyond Debate
The bench is a piece of debate prep. But the architecture it demonstrates is much more general.
Role specialisation beats role duplication. Giving four agents four different jobs produces better reasoning than giving four copies of the same agent the same question. The Opener’s constraints force it to define terms — something a general-purpose agent usually skips. The Closer’s constraints force it to weigh — something an agent will usually avoid unless compelled. Structure the roles and you force the reasoning.
Persona is a tool, not a gimmick. Injecting a rhetorical style into an agent isn’t cosplay — it’s a way of forcing the agent into a reasoning pattern it would not default to. A Buckley agent challenges premises before engaging arguments. A Wollstonecraft agent asks whether custom can survive rational scrutiny. A Malcolm X agent refuses the terms set by the opposition. These are reasoning moves disguised as voices.
Sequence matters as much as substance. An agentic system is not a pile of agents — it’s a pipeline. The Devil’s Advocate runs after the speeches because it needs to know who’s winning before it decides who to attack. The Closer runs last because it needs the full debate to weigh. The order of operations is architectural, not cosmetic.
Memory is a design decision. V1 used lossy round summaries. The new version maintains structured debate memory across turns — what’s been argued, what’s been conceded, what’s unanswered, what the floor is likely to raise. In any agentic system longer than a single exchange, the shape of your memory is the shape of your reasoning.
The output is the product. The debate itself is not the deliverable. The deliverable is a brief — structured, synthesised, usable in the chamber. This is true of every production agentic system: the chain of thought is a means; the distilled output is the end.
What Comes Next
The public tool above is Prep Mode — use it before the debate. Two further modes are in the V2 spec:
- Training Mode — drill specific debate skills (rebuttal under pressure, POIs, crystallisation) against an adversarial bench tuned to your weak points.
- Live Companion Mode — observe a real debate in progress. Track clash points, flag floor attacks as they emerge, produce a post-debate report that a speaker’s coach can actually use.
None of that is magic. It’s the same architectural pattern — specialised roles, distinct personas, structured memory, clear outputs — applied to three different use cases.
That’s the bigger point. Mixture of Agents isn’t a feature. It’s a way of thinking about intelligence as a system you design — not a tool you query. The interesting question has changed from what should I ask the AI? to how should I structure the reasoning process?
The Oxford Union debate tool is a small, concrete answer to that question. It’s free to use. It’s open to modify. It takes about ten minutes to understand end-to-end, and about two minutes to run.
Pick a motion. Stack your bench. Watch Hitchens and Thatcher fight about AGI.
Then ask what else this architecture could do.
Mark Preston is a technology executive and MBA graduate of Oxford University. He writes about AI, decision-making, and the future of intelligent systems. DeQuorum is the open toolkit for structured, adversarial AI reasoning.
