How to refine:
- Run your agent
- Observe the steps it takes to accomplish the task
- Note what works and what needs adjustment
- Use the Agent Builder again to describe changes, or edit the AOP directly
- Test and continue refining
Common refinements:
-
Adjusting edge case handling
- What happens when expected data is missing?
- How should the agent handle unexpected formats or values?
-
Clarifying which data fields to use
- Which specific column or field contains the information?
- What should the agent do if multiple fields could match?
-
Handling data variations:
- Different date formats
- Varying currency symbols
- Alternative naming conventions
-
Adding more approval checkpoints
- Where do you want to review before the agent proceeds?
- What threshold requires your input?
Conclusion
Each run is a training session that helps align your agent with your exact process. Use human-in-the-loop (HITL) approval points as strategic checkpoints to review the agent’s progress and identify where the AOP needs refinement. When making changes, focus on one modification per run—this makes it easy to understand the impact of each change and isolate what works. Refinement isn’t a sign of failure; it’s how you develop an agent that truly understands your unique business process.Things to know
The Agent Builder is looking at the current run that happened and is visible in the interface.Related
- Test Safely Before Going Live — Test plan template, mocking patterns, and environment separation before enabling autonomous Runs
- Guardrails for High-Risk Automations — How to classify risk and apply the right guardrails before an agent runs autonomously
- Designing Human-in-the-Loop Workflows — Using approval gates as checkpoints during refinement