A practical map of agent quality tooling

Updated 2026-10-12

AgentProbes helps application teams identify tools for a concrete quality problem: a release that changes answers, an agent that makes the wrong tool call, a retrieval step that returns irrelevant context, or an attack that changes intended behavior.

We describe documented capabilities and link the evidence so you can investigate a fit for your own stack. A directory entry is a starting point for testing, not a measurement of your application or a guarantee that a vendor prevents failures.

CUVISS maintains AgentProbes and also develops AI software. We disclose that connection and keep the editorial directory separate from any commercial showcase.

Our evidence and listing rules

• Classify the task. Entries are grouped by a documented use case. A category is an orientation aid; it does not imply a product lacks capabilities elsewhere.
• Keep the evidence attached. Capability notes show an official URL and a check date. We avoid performance promises and invite readers to inspect the source before an implementation decision.
• Keep submissions free. Editorial inclusion and correction are available without payment. A commercial discussion cannot buy a category position or rewrite a capability note.
• Disclose review relationships. A paid review, gifted account or product showcase must be disclosed where it appears. Vendor descriptions and editorial evidence serve different purposes.
• Correct the specific claim. Send the card name, disputed statement and replacement official evidence. We check the affected entry rather than treating a correction as a promotion request.
• Test on your own cases. An evaluation framework needs representative data, explicit criteria and a suitable environment. Directory inclusion cannot establish that a system is ready for production.