Skilly replaces unreliable LLM tool selection

Relying on LLMs to self-select tools is increasingly seen as a "prompting anti-pattern" due to high variance and hallucination risks.

• –Semantic classification via embeddings provides a deterministic confidence score that LLM reasoning lacks
• –Moving routing logic to an external layer like pgvector or a specialized embedding model (e.g., BGE-M3) can reduce prompt bloat by 60-80%
• –This approach is critical for "micro-agent" architectures where small local models (1B-3B) lack the reasoning depth for complex tool libraries
• –Future frameworks are likely to standardize on "Tool Search Tools" that retrieve schemas JIT rather than including them in every system prompt
• –Embedding-based routing also allows for better guardrails, as the system can physically block non-relevant tool definitions from ever reaching the LLM