Observability & FinOps

• Distributed tracing across model calls, tool executions, retrieval, memory and agent runs
• Correlation IDs that survive handoffs, queues and human approvals
• Time-travel debugging: replay of prompts, delegations and memory state at failure
• Quality and drift monitoring, including silent-failure detection
• Token metering, budgets, chargeback and cost per successful task
• Agent-versus-human attribution in every log

Accountable owner

SRE / Observability platform, with FinOps

• Head of SRE — accountable owner
• Observability engineer — instrumentation and conventions
• SRE — SLOs, alerting, on-call for agent estates
• FinOps analyst — budgets, chargeback, unit economics
• Data engineer — trace pipelines and retention

Ownership and phase are architectural recommendations, not assessed results.

Instrument from the first pilot request. Retrofitting correlation IDs after the fact is the most common and most expensive omission in this whole composition.

Signals to watch

• Trace completeness — share of tasks fully reconstructable
• Share of incidents replayable from telemetry alone
• Cost per successful task and budget burn rate
• Agent-versus-human attribution coverage in logs
• Mean time to detect a silent quality regression

Telemetry collection and aggregation may be tuned. Preserve outcome provenance, privacy, coverage and resource-limit alerts.

Failure modes

1. Classic APM is blind to the failures that matter here: infinite loops, hallucinated queries, silent quality decay
2. Roughly 47% of deployed agents are actively monitored; about 68% of organisations cannot distinguish agent from human activity in their logs
3. Missing correlation IDs, so an incident cannot be replayed or turned into a test
4. Token spend discovered at the invoice rather than in the budget

Existing tool and standards snapshot · approximately mid-2026. Examples, not required dependencies; verify current versions before adoption.

OpenTelemetry is the data plane only: it records, it does not judge. Pair it with the evaluation harness or you will have perfect traces of a system nobody can say is working.