Can classical agent architectures fix agentic AI accountability?

The paper argues that current agentic AI — systems that combine large language models with tool use, planning, and multi-step action — lacks the explicit architectures that the decades-old Autonomous Agents and Multi-Agent Systems (AAMAS) research tradition built for reasoning, communication, and governance, and that recovering those tools is necessary "to turn behavioural autonomy into responsible agency." It names four classical AAMAS resources agentic AI is missing: BDI (belief-desire-intention) architectures that give agents explicit mental states, formal communication protocols such as KQML and FIPA-ACL, mechanism design for aligning individual agent incentives with group objectives, and institutional or normative modeling for governance. The paper states that "AAMAS research developed rich, explicit models of cognition, communication, and social structure, but often lacked the flexibility and scalability that characterize current data-driven approaches," while "current agentic systems promise flexible reasoning and continuous action, but often lack the explicit architectures, communication and reasoning semantics, and normative grounding that earlier agent models provided."

Two properties of LLMs, in the paper's account, make this gap consequential. First, LLMs "imitate human intelligence in an input-output way" rather than executing a specifiable, verifiable function — so "the results from the past are no guarantee for results in the future," illustrated by an agent that books a convenient flight for a conference once and an unexpectedly inconvenient one the next time. Second, LLMs "live" in a world of data: patterns in training data, not an explicit specification of social context, determine their behavior, so performance degrades when context shifts or an exception arises. By contrast, BDI-style architectures give an agent explicit internal state and commitment mechanisms that support "introspection, communication of mental attitudes, and logical verification of consistency between goals and actions" — features the paper says LLM agents lack because their "intentionality is implicit, inferred from surface-level behavior rather than formally specified reasoning."

This extends the argument in What infrastructure must exist beyond training agents? that behavior training alone cannot supply governance — where that note locates the missing functions in external infrastructure built around the model, this paper locates them in classical AAMAS theory (protocols, mechanism design, institutional norms) that predates LLM agents entirely. It also sharpens How can users verify what an AI agent actually plans to do?: the complaint here is not just that users need a way to check an agent's plans, but that the agent itself has no internal commitment structure to check against, so verification has nothing stable to verify. It converges with Why do agents fail at identity verification and authorization? in calling for formal protocol-level governance, and cites the same NIST framework as a candidate alignment point for the benchmarks it proposes. And it gives a structural frame to Should AI systems stay collaborative rather than fully autonomous?: where that position paper argues from observed reliability failures, this one explains the failures as the predictable result of discarding AAMAS's explicit architectures.

The paper is a position piece, not an evaluation: it proposes that BDI architectures, formal protocols, mechanism design, and institutional modeling would make agentic AI more verifiable and accountable, but reports no implementation or test of any such hybrid system against an LLM-only baseline. Its examples — flight-booking agents, conference travel — are illustrative, not empirical. The implication the excerpt supports is a research agenda, to "operationalize these principles through concrete methods, tools, and governance structures," rather than a demonstrated fix; whether reintroducing classical agent formalisms is tractable at LLM scale remains open.