I saw OpenClaw's big claim about 99.9% reliable tool-calling for their agent framework. I'm always skeptical about these percentages from vendors, so I decided to test it myself before even starting a real integration.
I built a simple fuzzing script that feeds their API a thousand weird, malformed prompts and checks if the system still returns a properly structured JSON tool call or just breaks. I focused on edge cases: prompts with mismatched brackets, escaped quotes, or just plain gibberish mixed with a valid tool name. The key lesson was that their parser is robust, but I found a specific failure mode where a certain pattern of newlines and quotes would cause it to return an empty list instead of an error object. This doesn't mean the tool is bad, but now I know exactly where its boundary is, which is crucial for our customer support automation use case. I can build error handling around that specific quirk.