Preface
The Programme asks how multi‑agent systems can support strategic decisions whose tensions cannot be resolved by optimisation.
A strategic decision — to enter a market, to restructure an operation, to commit to a long technology bet — is unlike a benchmark task. Its inputs are uncertain; its stakeholders hold legitimate and irreducibly different interests; the question is not which option is correct but which option holds, under the disagreement that defines the problem. Half a century of work in management, strategy, and the philosophy of practical reasoning agrees on this: good strategic decisions preserve the tension that makes them strategic.
Contemporary multi‑agent LLM systems do almost the opposite. They are designed to converge — to debate until a single answer emerges, to aggregate until disagreement is smoothed away. They are evaluated on accuracy, on exact match, on the speed with which they produce a confident conclusion. For tasks whose answers are knowable, this is appropriate. For decisions whose answers are constructed, it is wrong.
The Programme develops dialectical synthesis as a distinct mode of multi‑agent reasoning: role‑conditioned agents produce extreme, internally consistent reframings of a decision; a Leader agent performs synthesis that preserves the irreducible tensions rather than collapsing them; the system produces not the answer but the conditions in which the decision‑maker can arrive at one. Evaluation is held to compute parity: any claim that the multi‑agent system does something a single agent cannot must hold under equal token budgets.
The work is conducted with strategy teams, executive committees, and operating institutions where the consequence of a decision is borne by the people who make it, and where the value of an AI counterpart is measured by what it leaves intact.
Horizon
The Programme takes a position.
The decisions that matter most inside an institution are the ones in which thoughtful people, in good faith, disagree. A system that resolves such disagreement on their behalf does not aid decision‑making; it bypasses it. It produces a confident answer, and in doing so it removes the very thing the decision was about.
Strategic Sovereignty is the property a team retains when its decision‑making capacity — its ability to see the tension, to name it, to hold it, and to choose under it — is left intact by the systems supporting it. A deliberative agent serves a team when it surfaces the disagreement that single‑answer systems suppress; when it makes legible the position that no individual member could have articulated alone; when it produces, at the end of the session, a decision that the team can defend as its own.
Tension as method is the working principle. Strategic Sovereignty is the horizon. The Programme refuses the trade in which a confident answer stands in for a held one.
Method
Four interleaved fronts — each grounded in the literature of strategic management, paradox theory, and multi‑agent systems, and disciplined by compute parity against single‑agent baselines.
The Programme takes its conceptual ground from paradox theory (Smith & Lewis 2011, 2022), the integrative‑thinking tradition (Riel & Martin 2017), and the broader strategic‑management literature in which the holding of tension — between exploration and exploitation, between stakeholder claims, between time horizons — is treated as a defining feature of strategic practice rather than a defect to be optimised away.
Theory enters the system as a concrete protocol. Role‑conditioned agents produce extreme reframings of a decision; a Leader agent performs dialectical synthesis through a structured grammar of tension‑resolution; outputs are returned with their constituent tensions preserved and named, not collapsed. Heterogeneity across role, model, and information‑sharing protocol is treated as a factor to be decomposed rather than averaged.
The Programme treats the question "does multi‑agent produce something single‑agent self‑consistency cannot?" as the standard against which any claim must be made — under equal token budgets, on the same prompts, with the same evaluators. Metrics include surprisal (Bayesian and counterfactual), tension‑holding (the proportion of stakeholder essences preserved across synthesis), and decision‑quality uplift as rated by expert practitioners. Multi‑LLM rotating judge panels eliminate self‑evaluation bias.
Strategic decisions are tested in the institutions that own them. The Programme partners with strategy functions, executive committees, and operating leaders, deploying the system within real cycles — KPI trees, decision cockpits, divergence labs, co‑creation sessions — and measuring its effect on the decision the team ends up making, the disagreement it ends up resolving, and the confidence with which it owns the outcome.
Now in progress
The Programme’s current anchor: the only rigorous way to ask whether multi‑agent reasoning earns its keep against a single careful model under equal compute — a methodological contribution the field has been waiting for.
Under equal compute, what does multi‑agent reasoning actually add over a single careful model? The Programme's lead contribution: the discipline that tells genuine multi‑agent value from compute spend.
Agents that disagree on purpose — preserving the irreducible disagreements that make a strategic decision strategic, instead of collapsing them into a smooth recommendation.
A position arguing that convergence‑optimised multi‑agent systems mis‑serve strategic decisions, and articulating dialectical synthesis as a distinct mode whose value emerges only when the comparison is made at equal compute.
Field protocols by which dialectical synthesis is operationalised within strategy functions, executive committees, and operating leadership — each protocol paired with a measurement framework specific to its decision cycle.
How the Programme is brought to bear
A worked sketch — not a client case — of how the Programme is brought to bear inside an actual strategic decision.
An executive committee is debating whether to fund a new platform business that would cannibalise a profitable existing line. LLM‑drafted memos keep producing a single confident recommendation that nobody in the room is willing to sign their name to.
The Programme begins with field work — not synthesis. What are the actual constituent positions in the room? What does each protect? Where do they touch a shared concern, and where does the disagreement become irreducible? The output is a structured map of the decision’s real geometry, not its surface framing.
Multi‑agent dialectical protocols are built — one per real stakeholder voice, each instrumented and pre‑registered. Validation is against a single careful model under equal compute, so the comparison is honest. Disagreements are held open by design; the agents are not permitted to vote.
The protocols are run inside actual decision cycles — quarterly strategy review, M&A diligence, capital allocation. Each cycle is paired with a measurement framework specific to its decision class. The deliverable is not a recommendation; it is a defendable position with the underlying tensions cited rather than hidden.
Not a smarter answer. A defendable one — and a strategy function that can say in the room why it did not collapse the disagreement that mattered.
Most Programme‑Ⅲ engagements run as Sponsored Research over 12–24 months, partnered with an executive sponsor inside the deciding institution.
Correspondence
For research collaborations, joint publications, advisory engagements, and field deployments —
A note via the form below — the Director responds personally.