Hi! Does anyone here have experience with Clay’s custom Claygent builder? I built a custom research agent designed to scan a company’s website and identify specific financial features. I defined each feature category very clearly, provided explicit examples for every variable, and during initial test runs the agent performed flawlessly. It correctly detected branded payments, accounting integrations, reporting, etc. But as we’ve started using it more, the accuracy and confidence have noticeably deteriorated. For instance, instead of identifying the branded payments product (like “XPay” or “Company Payments”), it has started outputting the underlying payment processor instead, which directly contradicts the instructions we gave it. It’s now ignoring parts of the prompt that it initially followed perfectly. We’re currently running it on the Argon model, so I’m also wondering whether switching to a different model might improve stability or reduce this kind of drift. If anyone has seen similar behavior with custom agents or has advice on prompt design, model selection, or keeping agents aligned with their instructions over time, I’d love any tips or guidance. Thanks!
This sounds like a classic prompt drift issue. When you see the same failure pattern repeatedly (like detecting payment processors instead of branded payment products), it's typically a prompt problem rather than a data issue.
For your model question - definitely try switching to GPT-5. It's now our default recommendation for complex web research tasks and shows 8-17% better research accuracy with more grounded results in source material.
To fix the drift, add specific instructions targeting your failure patterns. For example, explicitly state "identify the branded payment product name (like 'XPay' or 'Company Payments'), NOT the underlying payment processor."
Also make sure your prompts follow best practices: set clear context, provide step-by-step instructions, and be explicit about output format. You can use our metaprompters to help craft better prompts.
The key is systematic iteration - identify the pattern, update the prompt with specific handling instructions, then reprocess failed records through the updated prompt.
Is that what you were looking for?
If you still need help improving your custom Claygent builder or have more questions about prompt design and model options, I’m here to assist. Would you like to share more details about the challenges you’re facing or any specific areas where you want further support?
