Execution tracing for AI agents is substantially standardized in 2026 — decision accountability is not

Coverage and maturity matrix: what each layer defines, allows, or lacks

The left half of the matrix — transport, agent identity, messages, tool calls, retrieval, errors, evaluation — is green: execution can be reconstructed. The right half — intent, rationale, approval, post-condition — is missing in every layer. Current telemetry can say what model or tool was called, with which inputs, outputs, timing, and tokens; it usually cannot establish why a choice was made. Raw chain-of-thought is neither reliably exposed nor an appropriate universal primitive; useful decision observability needs concise decision summaries, alternatives considered, policy references, evidence, approvals, and outcome links instead.

A practical 2026 baseline while the standards converge

1. Adopt trace/span context and OTLP as the only transport; version every schema you emit.
2. Emit agent, inference, tool-execution, and retrieval spans using the official GenAI operation names or OpenInference span kinds.
3. Propagate session, task, and tool-call IDs across every span so runs can be stitched end to end.
4. Keep prompt and message payload capture opt-in and redacted — the official conventions already treat content as sensitive by default.
5. Add organization-level records the standards do not yet carry: decision summaries, policy and evidence references, approvals, state transitions, outcomes, and evaluation results, linked to trace IDs until vendor-neutral schemas mature.

Evidence rows behind this snapshot