TOTVS engineer Fabiane Nardon outlines strategies for preparing enterprise data for token-intensive AI agents. The approach balances deterministic transactional logic with non-deterministic LLMs by leveraging data mesh, low-latency database architectures, and semantic ontologies. Key tactics include dynamic MCP tool selection to optimize context windows and significantly reduce token overhead.
- Use semantic ontologies to map transactional data for efficient LLM context retrieval.
- Implement dynamic MCP tool selection to manage agent interactions and reduce latency.
- Balance deterministic logic with non-deterministic models to ensure security and precision.
- Optimize context windows to lower token costs in high-volume transactional systems.
