emilià

Building reliable AI agents for complex enterprise SaaS requires a 'harness' pattern with dynamic skill routing, deterministic tool execution, and analytical databases (DuckDB) to overcome LLM context limits and reasoning errors, rather than relying on raw model intelligence.

Embedding an in-process analytical database (DuckDB) allows AI agents to perform complex joins and aggregations on large datasets without overflowing context windows, solving the token cost and accuracy problems of naive LLM-based data processing.

The speaker argues that investing in tool harnesses, evaluation pipelines, and deterministic data layers (like DuckDB) delivers more reliable AI agents than chasing incremental model improvements, since current models are converging in capability for typical enterprise tasks.

Factorial's architecture shows how to deploy AI agents in regulated multi-tenant environments: enforce data isolation at the API layer, use deterministic tool proxies instead of raw model permissions, and price AI features via perceived-value credits rather than token consumption.