I’d start by enabling Performance Profiling in Alteryx to identify which tools are consuming the most time. With large datasets, joins, sorts, and complex formulas are often worth investigating first.

A few changes can make a noticeable difference:

• Use Filter and Select early to reduce unnecessary rows and columns.

• Check for duplicate join keys that might be multiplying records.

• Push joins and aggregations to the database using In-DB tools where possible.

• Avoid repeating expensive calculations, and consider caching intermediate results in .yxdb files.

I’d also check whether the workflow is hitting memory limits or spending too much time reading and writing temporary files.

Is the slowdown happening at a specific tool, or is the entire workflow running slowly?

I think one of the biggest reasons AI agents fail in production is that testing happens in controlled conditions, while real-world workflows are anything but predictable.

An agent might perform well on test data but struggle when APIs time out, customer information is outdated, or users ask questions outside its expected scope.

From a data perspective, I’d focus on three things:

1. Data reliability: Make sure agents access accurate, updated information with proper validation.

2. Monitoring: Track not just successful responses but failed tool calls, incorrect decisions, and unexpected behavior.

3. Human oversight: Set clear boundaries for when an agent should stop, ask for clarification, or escalate a decision.

I’d also recommend testing complete workflows rather than evaluating individual model responses.

The bigger challenge isn’t making an agent work once. It’s making sure it behaves reliably when conditions change.

How are teams currently measuring agent reliability after deployment, beyond basic accuracy and success rates?