An observation layer that scores your agents and flags what to correct

ElevenLabs launched ElevenAgents Spotlight, an observation layer that monitors, scores, and debugs conversational voice and text agents in real time.

ElevenLabs is adding an observation and improvement layer, ElevenAgents Spotlight, to its ElevenAgents. The idea is that once a conversational agent is in production, the real work begins, and the tool aims to drive continuous iteration by sifting through every conversation, both voice and text, in real time.

The system is built around a few key components. Each exchange is automatically grouped into themes and sub-themes, searchable by meaning rather than keywords, with a sentiment score assigned on the fly. You can then sort by volume, resolution rate, or sentiment to identify where to focus efforts, and then directly open the transcripts. Quality is defined in plain language: you write a criterion like "the agent recognized the customer's frustration before offering a solution," and each conversation is automatically evaluated against this benchmark, without manual review.

The tool continuously monitors success rates, latency, and evaluation results, flagging notable deviations—a drop in success rate or a spike in demand—to enable early reaction. It doesn't just identify issues; it also suggests actions, such as adding tests where none exist or activating a newer model, based on the agent's configuration and exchange history.

For integration, metrics, logs, and traces can be sent to Datadog, Grafana, or any OpenTelemetry-compatible backend, so agents appear in the same monitoring dashboard as the rest of the infrastructure. Access is available via the ElevenAgents documentation.