Argumentative Multi-Agent Planning

• Argumentative Multi-Agent Planning is a paradigm where agents synthesize plans through formal argumentation and defeasible reasoning under uncertainty.
• It integrates frameworks like AAF, CAF, and ABF to model coalition strategies and dynamic negotiation among multiple agents.
• Emerging approaches leverage logical, game-theoretic, and LLM-driven protocols to deliver robust, adaptive multi-agent coordination.

Argumentative Multi-Agent Planning (AMAP) refers to the synthesis of plans or strategies in multi-agent systems where agents coordinate, negotiate, or deliberate through formal argumentation mechanisms—sometimes in environments marked by uncertainty or partial information. This paradigm blends formal models of argument—such as Abstract Argumentation Frameworks (AAF), Control Argumentation Frameworks (CAF), and Assumption-Based Frameworks (ABF)—with multi-agent planning architectures and, in some applications, coalitional logics, game-theoretic semantics, or LLM-driven debate protocols. The result is a set of methodologies in which multi-agent planning, coordination, and robust reasoning under uncertainty are treated as structure-preserving argumentation tasks. Approaches to AMAP range from logic-based, strategic model-checking frameworks to LLM-enabled debate and persona-driven narrative generation.

1. Formal Foundations: Argumentation Frameworks and Multi-Agent Planning

AMAP systems are grounded on extensions or variations of the classical Abstract Argumentation Framework (AAF), $F = (A, D)$, where $A$ is a set of arguments and $D \subseteq A \times A$ is the attack relation. Multi-agent argumentative planning scenarios require extensions that encode (i) which arguments each agent can constructively deploy or control, (ii) the uncertain, dynamic evolution of arguments and attacks in a partially known environment, and (iii) the possibility of agent coalitions acting synchronously.

The Control Argumentation Framework (CAF), as defined by Dimopoulos et al. and advanced in "Modelling Control Arguments via Cooperation Logic in Unforeseen Scenarios" (Patil, 2022), distinguishes between fixed arguments ($A^F$), control arguments ($A^C$), and uncertain arguments ($\bar{A}$), together with their associated stratified attack relations. A CAF is defined as:

$$CAF = \bigl((A^F, R^F),\; (A^C, R^C),\; (\bar{A}, \bar{R} \cup R^{\leftrightarrow})\bigr)$$

with rigorous constraints on argument overlap and attack type. Control strategies are selections of $A^C$ designed to guarantee certain argument statuses regardless of which elements of $\bar{A}$ and $\bar{R}$ materialize.

Assumption-Based Argumentation Frameworks (ABF), central to Assumption-Based Planning (ABP) (Pellier et al., 2018), recast planning as a form of defeasible reasoning, where agents can propose conjectural plans ($\chi$), each annotated with explicit sets of assumptions (open subgoals). These can subsequently be defended, refuted, or discharged through dialogic exchanges, yielding a conjecture/refutation protocol that contracts or expands the plan until all assumptions are eliminated or deadlock is reached.

2. Coalition Logics and Strategic Semantics

AMAP requires a semantics for planning under coalition and adversarial action. "Modelling Control Arguments via Cooperation Logic in Unforeseen Scenarios" (Patil, 2022) generalizes ATL (Alternating-time Temporal Logic) semantics to CAFATL, embedding argumentation frameworks into multi-agent strategy synthesis. Here, a CAFATL model

$$S = (\Pi,\, \Sigma,\, Q,\, \pi,\, \Theta,\, \delta,\, K,\, \Upsilon)$$

combines states ($Q$), agents ($\Sigma$), available strategies per agent per state ($\Theta$), standard and argument-based environment transitions ($\delta$, $\Upsilon$), and a family of CAFs ($K$). The strategic modal language supports coalition operators

$\langle\!\langle C\rangle\!\rangle \varphi$

expressing that coalition $C$ can guarantee outcome $\varphi$ regardless of other agents' actions.

Strategy synthesis is realized through greatest fixed-point iteration on model states, seeking for each coalition $C$, the set of states from which they can enforce the relevant objectives. Monotonicity holds: enlarging the coalition can only increase the set of enforceable properties.

3. Dialogical and Defeasible Planning Protocols

Assumption-Based Planning (Pellier et al., 2018) operationalizes AMAP through a dialogic protocol centered on conjecture/refutation cycles. Each agent maintains local knowledge and planning operators. Plans may be proposed with unresolved assumptions, which are vetted by others. If an assumption cannot be discharged (proven or planned for) by any teammate, the plan is refuted; if defended, it is recursively refined.

The pseudocode for the multi-agent plan generation encapsulates this cycle:

function MultiAgentPlan(Agents, {E^j, Rules^j}, Goal G): χ ← Ag1.FindConjecture(E^1, Rules^1, G) H ← assumptions(χ) while H ≠ ∅: for each p in H: responders ← { Ag_k | p ∉ E^k } if responders == ∅: H ← H \ {p}; continue for each Ag_k in responders: ψ ← Ag_k.FindConjecture(E^k, Rules^k, p) if ψ ≠ FAILURE: χ ← mergePlans(χ, p, ψ) H ← assumptions(χ) break else: χ_alt ← nextAlternativePlan() if χ_alt == FAILURE: return FAILURE χ ← χ_alt H ← assumptions(χ) break return χ

This dialogic approach tightly couples task decomposition, allocation, and conflict resolution, supporting plan construction even under partial or discordant agent knowledge.

4. Dynamic Environments and Uncertainty

AMAP frameworks explicitly address both environmental and argumentative dynamics. In the CAFATL setting (Patil, 2022), every global step is a pairwise update: the world state evolves according to the joint action, and the environment's argumentation structure (the CAF) can be updated via $\Upsilon$ to reflect the appearance or disappearance of arguments and attacks—these may represent unforeseen contingencies. This enables agents and coalitions to synthesize robust strategies that incorporate the possibility of argumentation structure changes and runtime emergence of uncertainty.

In ABP (Pellier et al., 2018), missing preconditions, contradictory literals, or numeric constraints are all modeled as assumptions, to be validated or discharged during planning. The exchange of conjectures and the systematic pruning of risky plan branches via branch-and-bound ensures minimal-risk plans with explicit accommodation of knowledge gaps. Termination of the protocol is assured under bounded assumption depth and operator space.

5. Data-Driven and LLM-Based AMAP: Persona-Driven Planning

Recent approaches (Hu et al., 2024) integrate data-driven, LLM-based agents into AMAP, leveraging persona assignment and debate-driven protocols for argument generation and text planning. The Debate-to-Write framework structures the planning process as a multi-agent debate, where each agent is instantiated with a distinct persona and high-level claim, enabling nonlinear refinement of argumentative structure.

Planning proceeds through iterative, round-based exchange:

1. Assign personas (symbolic profiles).
2. Conduct an agent debate, including adversarial critique.
3. Extract a structured plan (list of main points and subpoints) from the debate history.
4. Generate the final argumentative text in accordance with the plan.

The process is summarized:

$$(p_i) \rightarrow debate(D) \rightarrow plan(P) \rightarrow essay(y)$$

This debate protocol produces outputs with empirically validated gains in both argumentative diversity and persuasiveness, as assessed by GPT-based measures and human evaluators. Diversity is measured via Self-BLEU, embedding-based similarity, and perspective diversity metrics.

6. Complexity, Properties, and Comparative Perspectives

Model checking in logical frameworks such as CAFATL is polynomial in model size and linear in formula length, mirroring standard ATL results (Patil, 2022). The fixed-point iteration for strategy computation converges in at most $|Q|$ steps. In ABP, search complexity is exponential in plan length and number of assumptions, comparable to hierarchical task networks with added assumption-layer branching; branch-and-bound heuristics and depth bounds are used for practical restriction (Pellier et al., 2018).

AMAP contrasts with classic multi-agent planning by tightly integrating argumentation for plan construction and validation, avoiding early hard commitments or purely negotiation-based coordination. The explicit management of assumptions, defeasible rules, and dynamic coalitions yields greater robustness to partial knowledge and unforeseen events.

7. Applications and Outlook

Argumentative Multi-Agent Planning is applied to domains where robust, explainable coalition strategies under uncertainty are critical. This includes multi-agent orchestration in dynamic, open environments, collaborative text generation under diverse perspectives (Hu et al., 2024), and planning with partial or conflicting information. Ongoing research explores time-bounded strategic operators, epistemic modalities to track distributed agent knowledge, and integration with preference annotation on arguments (Patil, 2022).

A plausible implication is that AMAP provides a framework for explainable, adaptive, and coalition-capable planning in both symbolic and neural multi-agent systems. This enables planning, negotiation, and dynamic allocation of subgoals in uncertain and adversarial settings, tightly coupling deliberative and argumentative reasoning layers.