Hi Product Hunt 👋 When building agents the same thing kept breaking. Not the reasoning, the execution. The model picks the right tool, then sends a payload with a hallucinated field, or hits a 429, or catches a five minute outage at the vendor. The agent has no principled way to react, so it retries a 400 forever, or re-prompts itself on a 503 and mangles a payload that was already correct. klanex is the layer I wanted. You POST a tool-use intent and get an execution ID in milliseconds. It validates the payload against your JSON Schema before anything leaves the building, then owns the retries, backoff, circuit breakers and idempotency, keeps your API keys KMS-sealed and out of the model's context, and reports the outcome over a signed webhook with an audit trail you can replay byte for byte. The part I most want feedback on is error design. Failures come back classified, and the ones a model can actually fix carry an llm_hint written to be pasted straight into its context. Transient and auth failures explicitly tell the model not to touch the arguments, which is what stops a good payload getting wrecked on retry. You can try it right now without an account: curl -sX POST https://api.sandbox.klanexai.com... That mints a disposable sandbox key. 100 executions, no signup, no card. The error taxonomy is open source and stands alone too, Apache-2.0 and zero deps: github.com/chrassy/agenterr How are you handling failed tool calls today? I'm especially curious where a four-way split (retry / re-prompt / escalate / human) breaks down for your tools.

klanex

Reliability layer for AI agent tool calls

Reliability layer for AI agent tool calls