Vinay Goel

More than half of global companies already have voice agents answering their phones. The hard part isn't answering. It's knowing which calls cost you a customer. Two callers ask the same question. One gets an answer in 0.6s. The other sits through 4.4s of silence, says "okay, forget it," and hangs up. He cancels his order 3 minutes later. The call log says both went fine. Agent Analytics now does voice. Every call becomes Amplitude events on the same user ID as the rest of your product, so you can: - See the problem. Open any call and find the slow tool behind the silence. - Score it. Built-in signals and your own evals run on every call. - Tie it to the business. Put the wait next to the cancel, the churn, the repeat order. - Ship the fix, and see if it worked for your callers. Stop shipping on vibes.

Fine-tuning on raw prod traces teaches the model your agent's failure modes too. Prod logs are full of sessions where the agent looped on a tool call, the user rephrased the same question three times, or gave up halfway. Train on all of it and the model learns those habits. How we handle this in @Amplitude_HQ Agent Analytics, for agents running on @FireworksAI_HQ: - Each response keeps its Fireworks request ID, so a label maps back to the exact inference - Built-in signals score every session (task completion, friction, negative feedback), each with a written rationale - Your own evaluators (LLM, code or decision models) run on top, calibrated against sessions you've reviewed - Export passing sessions for SFT, or pass and fail sessions, labeled, for DPO / reward models - A/B the tuned model against prod by experiment arm: quality, cost and latency side by side You choose the training set and run the export yourself, so your data goes only where you send it. The result: a tuned model you can check against prod on your own traffic before you switch. How are you choosing training data from prod today?

Agent Analytics meets you where you are. Where your agent runs: - Python or Node: wrap the client with the SDK. OpenAI, Anthropic, Gemini, Bedrock, Azure OpenAI, Mistral. - Frameworks: LangChain, LlamaIndex, CrewAI, OpenAI Agents SDK, Claude Agent SDK, Vercel AI SDK, Mastra, Pydantic AI, AWS Strands, Google ADK, Microsoft Agent Framework. - Already on OpenTelemetry: add Amplitude as a second exporter. OpenLIT, Traceloop, Bedrock AgentCore, and OpenInference (Agno, smolagents, DSPy, Haystack, AutoGen) work as they are. - Gateways and local models: LiteLLM, OpenRouter, Ollama. - Anything else (browser, edge, Java, Go, Ruby, Cursor Agent SDK): post [Agent] events to the HTTP API. Where your traces live: - LangSmith, Langfuse, Braintrust: a small job forwards settled threads. - Sierra, Decagon, Fin, Agentforce: same pattern, pull or push. - Snowflake, BigQuery, Databricks, including MLflow traces: Amplitude imports them with your SQL on your schedule. Agent Analytics turns agent traces and evals into product events, so you can tie agent quality to conversion, retention, and revenue. The recommended setup is one prompt to your coding agent. It does the wiring and stops to ask what it can't infer, like which field holds your user ID. Don't see your stack? Reply with what you run.