Agentic Business Process Management: A Research Manifesto

Manifesto paper authored by 18 researchers from the BPM, multi-agent systems, and AI communities, derived from Dagstuhl Seminar #25192 "AUTOBIZ" and the PMAI'25 workshop at ECAI 2025. It articulates the conceptual foundations of Agentic Business Process Management (APM) — an extension of BPM for governing autonomous agents.

Summary

The manifesto argues that classical BPM/RPA systems assume predefined workflows executed by passive software components, but the emergence of LLM-based agents demands a new paradigm where software and human agents act as primary functional entities that perceive, reason, and act within explicit process frames. This shift is characterised not by eliminating control but by framing it: agents retain autonomy yet remain aligned with organisational goals through normative constraints.

The paper introduces a two-level conceptual architecture:

Macro level (APM System): the socio-technical environment hosting interoperating agents, tools, and protocols (MCP, ACP, RDF, KQML), where a framing mechanism imposes process-awareness and goal-alignment on agents.
Micro level (APM Agent): individual agents running a Perceive–Reason–Act loop with three conceptual modules (perception, reasoning, action), drawing on foundational ontologies (DOLCE, GFO), agent-oriented modelling (i*, Tropos), and classical AI constructs (BDI, FIPA).

Four capabilities are required of an APM agent, ordered intentionally:

Framing — internalising normative frames so the agent is process-aware and guard-railed.
Explainability — articulating the rationale behind decisions; a prerequisite for trust, debugging, and regulatory compliance (GDPR, EU AI Act).
Conversational Actionability — interacting with humans and other agents via natural language or formal protocols, and mapping conversations to four enactment roles (Query / Recommend / Create / Execute).
Self-Modification — adapting (short-term, instance-specific) and evolving (long-term, model-level) based on experience.

A research agenda enumerates 19 named challenges across these capabilities (F1–F4, X1–X5, A1–A5, M1–M5) plus 5 cross-cutting challenges (C1–C5): legacy onboarding via agent-centric process mining, security and prompt-injection risk, benchmark contamination, liability under the EU AI Act, and engineering methods.

Framing distinctions introduced

Normative frame vs operational frame. Normative frames specify deontic requirements (obligations, permissions, prohibitions); operational frames (BPMN, DECLARE) specify how to act. APM treats the normative as first-class.
Autonomy ≠ automation. Automation executes predefined tasks exactly; autonomy perceives, reasons, and chooses within a frame.
Adaptation (ephemeral, instance-level) vs evolution (persistent, model-level) — see concepts/self-modification.
Centralised vs distributed intelligence — affects where frames are placed (on process vs on individual decision-makers).

Positioning

Successor-in-spirit to the Process Mining Manifesto (2012).
Makes agent-centricity explicit where the AI-Augmented BPMS manifesto (2023) left it implicit.
Complements Large Process Models (2024) while asserting LLMs alone are insufficient.
Parallels the four-level analytics pyramid of Chapela-Campa & Dumas (2023).

Connections

Concepts: concepts/agentic-bpm, concepts/framed-autonomy, concepts/process-awareness, concepts/explainability-apm, concepts/conversational-actionability, concepts/self-modification, concepts/perceive-reason-act

Authors (entities): entities/diego-calvanese, entities/angelo-casciani, entities/giuseppe-de-giacomo, entities/marlon-dumas, entities/fabiana-fournier, entities/timotheus-kampik, entities/emanuele-la-malfa, entities/lior-limonad, entities/andrea-marrella, entities/andreas-metzger, entities/marco-montali, entities/daniel-amyot, entities/peter-fettke, entities/artem-polyvyanyy, entities/stefanie-rinderle-ma, entities/sebastian-sardina, entities/niek-tax, entities/barbara-weber

Frameworks/methods: frameworks/bdi-agents, frameworks/fipa, frameworks/tropos-i-star, methods/agent-oriented-process-mining

Related sources (stubs): sources/2023-dumas-ai-augmented-bpms, sources/2012-vanderaalst-process-mining-manifesto, sources/2024-kampik-large-process-models, sources/2023-chapela-campa-augmented-process-execution

Open questions raised by the source

How to operationalise frames on real-world symbolic data (F3)
How to design incentive mechanisms that make frame-compliance intrinsically valuable to agents (F4)
How to balance autonomy / delegation / control at the KPI level (A4–A5)
Benchmark contamination risk (C3) — a methodological warning relevant to any future PPM evaluation work.