Langfuse
Records how the agent ran: traces, prompts, latency, token cost, and evaluation scores. Engineers use it to debug and improve the agent.
Integrations, Langfuse
Langfuse tells engineers how the agent behaved: which model it called, how long it took, what it cost, and how it scored. VeriProof decides whether each business action is allowed, sends exceptions to the person who can approve them, and keeps the record risk and audit rely on. The two can share one trace, so nobody argues over whose version of events is right.
A refund agent's completion rate drops on Tuesday afternoon, and two people want to know why.
The engineer opens Langfuse. Latency and model cost look normal, but the trace for each stalled refund ends in a VeriProof span with the policy result escalate and the reason the rule gave. Nothing is broken; refunds are waiting for a person.
The operations manager who owns the agent opens VeriProof. A rule change approved that morning lowered the refund limit for new accounts, so 40 refunds went to supervisors instead of completing on their own. She can see who approved the change and decide whether the new limit is worth the extra review work.
Same trace, same identifiers. Each person answered their own question without asking the other to pull records.
Records how the agent ran: traces, prompts, latency, token cost, and evaluation scores. Engineers use it to debug and improve the agent.
Checks each business action against your rules before it runs, routes exceptions to the person with authority, and keeps a decision record for risk, compliance, and audit.
Langfuse shows what happened after the fact. The VeriProof check happens before the action, through the SDK, so it can stop or pause the work.
The VeriProof core SDKs for Python and TypeScript write a span for every checked action through your application's own OpenTelemetry setup. The span carries the action name, the policy result, the reason, and the identifier that opens the matching VeriProof record. It carries no prompt or response content.
If your application already sends its traces to Langfuse, that span appears there next to the model calls with no extra configuration. There is no separate Langfuse package to install.
VeriProof also accepts standard OpenTelemetry traces over its API, so one OpenTelemetry provider can export to both destinations when you want VeriProof to hold the trace as well.
No. Keep your Langfuse setup as it is. The VeriProof span travels through the OpenTelemetry provider your application already uses.
No. The span holds the action name, the policy result, the reason, and identifiers. Your application chooses what else it sends to VeriProof, and can redact sensitive fields before anything leaves it.
No. Langfuse is built for debugging and evaluation. The VeriProof record holds the rule version, the reviewer's decision, and the reported outcome, and compliance and audit read it in their own portals.