A trace is a vocabulary, not a verdict

An agent crosses many systems in a single piece of work. A client invokes a workflow, a router selects a model, the model requests a tool, the tool calls a service, and a second agent finishes a subtask. Without shared telemetry, each boundary invents its own names. The resulting traces are difficult to join and almost impossible to compare.

OpenTelemetry semantic conventions exist to give operations, attributes, metrics, and events common meaning across libraries and platforms. Its GenAI work now covers agent invocation, tool execution, model usage, evaluation events, and MCP-oriented observations. This is important infrastructure for interoperable evidence.

It is also easy to overread. A trace describes what instrumented components emitted. It does not determine whether the action was authorized, whether the result was fair, whether the user understood a confirmation, or whether an omitted span hides the decisive step. Shared vocabulary improves the account. It does not complete it.

Common names reveal common gaps

A consistent operation name such as agent invocation or tool execution lets operators search across frameworks. Agent and model attributes can identify which configured system participated. Tool arguments and results can make a failure reproducible. Once many systems use the same fields, dashboards and evaluations become portable.

The current OpenTelemetry GenAI semantic conventions are explicit that some content-bearing attributes may contain sensitive information. Arguments, results, system instructions, and messages can expose credentials, personal data, proprietary text, or high-cardinality values. More complete capture can improve reconstruction while creating a second uncontrolled copy of the very material the system was meant to protect.

That tradeoff is not solved by one global logging switch. Different operations need different evidence. A payment tool may require normalized amount, destination, authority, and response code while excluding the model's surrounding conversation. A retrieval operation may need dataset version and query digest while retaining returned content in a governed source repository. Telemetry should point to controlled evidence when copying it would create unnecessary risk.

A semantic convention tells systems how to name an observation. It does not tell institutions what the observation proves.

Common names also reveal absence. If one framework emits tool-call spans and another emits only model calls, an end-to-end trace may look continuous while losing the moment of action. Versioned conventions help identify this gap because the expected operation vocabulary is known. Missing instrumentation then becomes a visible limitation rather than an invisible assumption.

Causation needs more than correlation

Trace and span identifiers connect operations into a tree or graph. They show temporal and causal relationships asserted by instrumentation. In distributed agent work, those relationships can break at queues, webhooks, retries, and delegated tasks. A fresh trace may continue the same institutional action. One trace may also contain several proposed actions of which only one was authorized.

The record should therefore carry stable action and task identifiers alongside telemetry correlation. A retry should point to the attempt it repeats. A delegated task should identify the handoff that caused it. A webhook should identify the task event it reports. These links let reviewers distinguish technical trace structure from the institution's account of one consequential act.

Versioning matters here. The general OpenTelemetry specification documents stability levels and version selection because semantic meanings evolve. The version-selection guidance acknowledges that instrumentations may emit different convention versions. A query written against today's attribute names can silently miss yesterday's data unless the evidence records which schema applied.

Clock assumptions should also remain explicit. Distributed spans use timestamps from different hosts and may be exported late or sampled. Their order can be helpful without being courtroom chronology. Sequence identifiers, service acknowledgements, and signed timestamps may be needed when exact ordering determines responsibility.

Build a record from observations

A durable account can use telemetry as one source. It selects the observations relevant to the action, adds policy and authority context, references controlled inputs and outputs, records known gaps, and preserves the semantic-convention version. The resulting record is smaller than a full trace in some places and richer in others.

Sampling policies belong in that account. Head sampling may discard a trace before an error becomes visible. Tail sampling may retain failures while losing the ordinary population needed for comparison. Vendor pipelines can redact, aggregate, or drop attributes under load. A reviewer should know whether the trace is complete by design, selected because it looked unusual, or reconstructed from signals with different retention periods.

Retention also changes meaning over time. Metrics may survive after spans expire; a record may preserve a digest after raw tool content is deleted. This can be correct, but the transition should be deliberate. Evidence references should state the retention class and expected expiry so a future reviewer does not discover that the decisive observation vanished while its dashboard aggregate remained.

This separation protects both functions. Observability remains optimized for operating systems: diagnosis, performance, trends, and correlation. Decision evidence remains optimized for custody and review: completeness relative to an obligation, stable meaning, controlled disclosure, and an accountable owner.

The two systems should be linked. A reviewer should be able to move from an action record to relevant spans, and an operator investigating a span should be able to identify the consequential action it served. Neither should silently substitute for the other.

Shared telemetry is one of the most useful developments in agent interoperability because it makes systems observable in a common language. The final discipline is grammatical: do not mistake a well-formed description of activity for a verdict about what that activity meant.

— Dispatches · Summit Cognitive