What is an argumentation framework?

An argumentation framework is a formal structure made of arguments, that is claims with a reason behind them, tied by a single kind of relation, attack: if A attacks B, then B is acceptable only if A in turn does not hold up. It answers, in a checkable way, a question that now comes up whenever several agents built on LLMs produce incompatible claims: which one prevails, and on what grounds. The framework supplies a criterion declared up front, instead of an implicit choice made downstream. The model goes back to Phan Minh Dung, who presented it at IJCAI in 1993 and published it in Artificial Intelligence in 1995 as the basis of nonmonotonic reasoning, the kind where a conclusion already reached is withdrawn once new information arrives. On that minimal structure you compute acceptability semantics: groups of arguments that do not contradict each other and in which every member is defended against outside attacks, the formal answer to the question "which claims survive".

See also: Agent Swarm · A decentralized network of AI agents that coordinate without a central controller, each reacting and communicating with peers toward a shared goal.

The problem it solves: a majority is not a criterion

When three agents answer the same question differently, the conflict is settled today by majority vote or by keeping the last answer. A majority rewards redundancy, since three agents reading the same source count as three, and the last answer rewards execution order. Neither leaves a trace explaining, six months later, why one claim won. A graph of arguments and attacks gives two distinct things instead: an explicit criterion for which sets of claims are coherent and defensible, and the structure of the disagreement, which stays inspectable and is the natural bridge to explainable AI and to any duty to justify a decision. The distinction from the reflection pattern is sharp: there a single agent critiques and refines its own output, here the conflict is between different agents and neither can arbitrate on its own. Coordination among agents answers the question "who does what"; argumentation answers "who is right".

The technique is not ready to use yet, because of two limits. The first is building the graph: extracting arguments and attacks from natural language text remains the hard, unsolved part, and a misidentified attack yields a formally impeccable verdict on wrong premises. The second is cost, which varies with the chosen semantics: the most cautious one, the one the example below effectively uses, is computable in polynomial time, that is at a cost that stays manageable even on large graphs; the ones that turn prohibitive are the semantics admitting several alternative verdicts.

An enterprise example

In a supplier onboarding pipeline, three agents assess whether to grant sixty-day payment terms. The agent reading contracts concludes yes, because the signed draft sets that term. The agent connected to the ERP concludes no, because the same supplier has three missed payments in the last twelve months. A third agent finds a company filing from two weeks ago showing the missed payments settled. A majority vote has no use for it: it counts two risk signals against one favorable and closes with a no. Modeled as attacks, the picture changes: the missed-payments argument attacks the contractual one, but the recent filing in turn attacks the missed-payments argument, which therefore fails to defend itself, and the yes becomes acceptable again. The result is not "the AI decided", it is a chain the credit manager can read and contest link by link.

Why it matters for decision makers

Anyone moving agents into production on decisions with consequences has to be able to explain how the system chose, whether it erred or not. Recent research points that way, both by training models to resolve conflicts with abstract argumentation and by representing disagreement among LLM agents as a graph of weighted supports and attacks, as in a 2025 multi-agent forecasting framework built on quantitative bipolar argumentation, alongside the original paper by Dung. Before funding a project like this, ask whether conflict between agents in the process is already observable, because without it the framework has nothing to order.