Human Judgment in the Agent Improvement Loop

AI agents work best when they reflect the knowledge and judgment a team has accumulated. Documented knowledge matters, but the key is transferring the Tacit Knowledge inside employees' heads to the agent. This requires an Improvement Loop that includes input from domain experts.

For example, when building a Copilot for traders, the agent must have both financial domain knowledge (e.g., what "today's exposure" means) and technical database knowledge (e.g., which tables are reliable). To achieve this, teams need to work with relevant experts to teach the agent context that isn't written down anywhere.

Human involvement improves each component of the agent as follows.

• Workflow Design: Rather than relying solely on the LLM's autonomy, Deterministic Code can be used at steps requiring regulation or risk management to strictly control execution order and add verification processes.
• Tool Design: The names, parameters, and descriptions of the tools the agent uses must be optimized. Considering the trade-off between flexibility and control, teams should configure a toolset that fits business constraints and verify performance and risk through Evaluation.
• Agent Context: Beyond a simple system prompt, providing rich context from the start of execution improves the agent's accuracy.