Autonomous Skill Development: Neuro-Symbolic Agents that Self-Generate Verifiable Tools
| Publicatietype: | In proceedings |
| Citatie: | |
| Publication status: | Published |
| Boektitel: | 24th International Conference of the Italian Association for Artificial Intelligence |
| Deel: | LCNS |
| Jaar: | 2026 |
| Maand: | Oktober |
| Uitgever: | Springer |
| Locatie: | Perugia |
| Samenvatting: | Large language model (LLM)-based agents can address highly heterogeneous tasks. However, their result and justifications are generated through opaque (non-inspectable) sub-symbolic processing. Thus, even a plausible explanation may be post-hoc and unfaithful to the process that produced the answer. Such behavior undermines safety, auditability, and accountability in high-stakes domains such as healthcare and finance. This paper presents Autonomous Skill Development (ASD), a neuro-symbolic agent architecture that separates open-ended language reasoning from deterministic computation. Rather than relying on the LLM to repeatedly solve and explain computational problems, ASD dynamically generates, validates, stores, and reuses explicit symbolic skills. A neuro-symbolic router first determines whether a request contains a deterministic computation that can be extracted and formalized. When such a computation is identified, ASD generates a skill as a pure function exposed through an interoperable Model Context Protocol (MCP) tool. The skill is then evaluated through a layered validation pipeline that combines deterministic tests with an independent LLM judge and, once validated, is stored in a reusable, versioned skill library. Subsequent computations are performed by inspectable code, while the LLM is restricted to interpreting and communicating the resulting output. Each answer is therefore accompanied by explicit provenance identifying the skill executed, the input values provided, and the result produced. Unlike a free-form LLM justification, this trace is faithful by construction because it directly reflects the computation that generated the answer. Dynamic skill generation is essential for scalability, as manually implementing rules for every possible deterministic subtask would be impractical. We evaluate ASD in the nutrition domain by examining the skills it generates and validates, and assessing whether delegating deterministic computation to verified tools allows smaller language models to retain correctness. |
| Trefwoorden: | Agent Skills, Agentic AI, LLM-based Agents, MCP Tools, Neuro-Symbolic AI |
| Auteurs | |
| Toegevoegd door: | [] |
| Totaalscore: | 0 |
|
Bestanden
|
|
|
|
|
Aantekeningen
|
|
|
|
|
|
Onderwerpen
|
|
|
|
|
