From tasks to autonomous, goal-driven agents
This work proposes an agentic AI paradigm that replaces rigid task sequences with autonomous, goal-driven agents orchestrated around business objects. By formalizing goals, objects, and agents—and how they split, merge, and precede one another—the approach enables flexible, context-aware automation in real time. It also confronts governance head-on, urging transparency, oversight, and regulation as agent autonomy scales across enterprise workflows.
Points clés
- The paper replaces traditional task-based workflows with an agentic AI model centered on goals, objects, and autonomous agents for dynamic industrial environments.
- Agents use CRUDA capabilities (create, read, update, delete, archive), activating on trigger objects and releasing final objects upon achieving their goals.
- Goals are sequenced via a precedence relation and support split (AND/OR/XOR) and merge constructs, enabling non-deterministic, parallel, and adaptive execution paths.
- A pizza delivery example (agents a1–a5) demonstrates end-to-end flow—from order acquisition to delivery—mapping trigger/final objects and agent competences to goals.
- The formal agent model is a 6-tuple (aID, Ca, OTa, ORa, OFa, ga), while goals are (gID, Og, Ag); merge goals compose objects as OG = ⋃ OF_i.
- An agent-based business process (ABP) is defined as (OS, OE, OR, G, C, A), enabling verification of activation conditions, redundancy checks, and execution ordering.
- The approach leverages LLMs and generative AI for autonomous planning, reasoning, memory, and context-aware decision-making in enterprise processes.
- Governance and risk are emphasized, calling for transparency, auditability, human oversight, and regulatory frameworks to align agent behavior with human intent.
- Related research includes EvoFlow (Zhang et al.), modular agent architectures for parallel execution (Niu et al.), and compound AI for enterprise orchestration (Kandogan et al.).
- Authors are Mohammad Azarijafari and Luisa Mich (University of Trento) and Michele Missikoff (CNR IASI); presented at Ital-IA 2025 (CINI, Trieste, June 23–24, 2025) under a CC BY 4.0 license.
À retenir
Start small: pick one messy process, define clear goals and objects, and let a pilot agent handle the boring bits—while you keep a very human hand on the off switch. Document triggers and outputs (yes, all of them), add monitoring like you mean it, and rehearse your escalation plan before a chatbot invents a pizza. Finally, involve compliance early; it’s cheaper than apologizing to your auditor later—with or without a slice.
Sources
Quiz sur le document: 10 questions
