Toward Intelligent Skill Discovery in A2A via Fine-Tuned Small Language Models
| Type of publication: | Inproceedings |
| Citation: | |
| Publication status: | Accepted |
| Booktitle: | 24th International Conference of the Italian Association for Artificial Intelligence |
| Series: | LNCS |
| Year: | 2026 |
| Month: | October |
| Publisher: | Springer |
| Location: | Perugia (Italy) |
| Organization: | AIxIA |
| Abstract: | The Agent-to-Agent (A2A) protocol standardizes communication among heterogeneous LLM-based agents through Agent Cards, task lifecycle management, and bindings. However, while Agent Cards describe agent capabilities, A2A does not yet define a canonical registry and search layer for a seamless semantic discovery of the agents' capabilities across distributed or consortium-based catalogs. As a result, discovery is left to ad hoc registries and implementation-specific mechanisms, making it difficult to provide reliable matchmaking across heterogeneous agent ecosystems. The closest historical analogy, the FIPA Directory Facilitator, addressed a similar need but relied on formal service descriptions and shared ontologies that are difficult to scale to open-ended LLM-agent environments. This position paper argues that A2A requires a modern counterpart of the Directory Facilitator: a scalable and auditable discovery layer that maps natural-language intents to machine-readable Agent Cards. The proposed approach treats A2A agent and skill discovery as a semantic retrieval, re-ranking, and skill-selection task. It combines bi-encoder retrieval for efficient candidate selection with a fine-tuned Small Language Model (SLM) that re-ranks the retrieved Agent Cards, selects the skills supporting each match, and generates human-readable explanations. The paper further outlines an open research agenda for reproducible A2A discovery, including public benchmark construction, realistic query design, lexical and neural baselines, and evaluation protocols combining classical information retrieval metrics with LLM-assisted evaluation of semantic relevance, skill selection, and explanation quality. We argue that such a framework is needed to move A2A discovery beyond operational registries, and beyond the small sub-consortia in which agents currently interoperate, toward experimentally grounded semantic matchmaking. |
| Keywords: | A2A Registries, Agentic AI, Directory Facilitator, MAS, Skill Discovery, small language models |
| Authors | |
| Added by: | [] |
| Total mark: | 0 |
|
Attachments
|
|
|
|
|
Notes
|
|
|
|
|
|
Topics
|
|
|
|
|
