Where do your AI agents repeat expensive validation that probably shouldn't be repeated?

I’m building SeenRelay around a simple question:

When an agent checks something it has already checked recently, when should it trust prior evidence — and when should it always go back to the source?

Examples I’m especially interested in:

• repeated paid web searches

• browser or portal checks

• metered extraction

• MCP/API read-only calls

• multiple agents independently validating the same bounded fact

The obvious answer is “cache it,” but that breaks down quickly when freshness matters, when providers already have native caching, or when the underlying fact can change.

My current rule is deliberately conservative: measure first, prefer native/source controls first, and only reuse when the real workload demonstrates that it is both safe and economically worthwhile.

I’d be interested in concrete examples from people building coding agents or agent workflows:

What validation do your agents repeat today that feels wasteful — and what prevents you from simply caching it?