Abstract

A method for emitting structured observability telemetry during AI agent execution by keying each telemetry span to entries within a skill instruction file. Each span record carries a span identifier, a parent span identifier for trace hierarchy, a reference to the originating skill file, a reference to the specific capability entry being exercised, an optional reference to any constraint that was evaluated during the span, a hash of the pre-execution context, a hash of the post-execution output, elapsed duration, a list of tool invocations made during the span, and an optional score field that records a numeric value produced by a separate output evaluation step if one occurred. Spans are emitted over OpenTelemetry-compatible transports, enabling ingestion by existing observability backends such as Datadog, Honeycomb, Jaeger, and Grafana Tempo without custom adapters. The schema performs no enforcement; it is a recording format only. Trace replay at debug time is enabled by pairing the emitted hashes with a content-addressable store that resolves hashes back to full context and output payloads. A trace query language extension allows operators to filter spans by capability entry identifier, enabling inspection of all invocations of a specific capability across a fleet of agents.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS