Agentic Workflows Overview

Agent workflows create harder spend problems than single chat calls because one user request can trigger many model calls. A planner may call a model, spawn researchers, retry failures, fall back providers, ask a reviewer, then discard some branches.

Blackridge can record model-call economics for those calls. It can reconstruct workflow cost best when each request carries stable metadata:

Metadata	Why it matters
Tenant/app/team/user	Ownership and chargeback.
Workflow ID	Groups related model calls.
Workflow run ID	Separates one execution from another.
Root request ID	Connects all child requests to the originating call.
Parent request ID	Builds lineage edges.
Agent ID and step/task ID	Explains which agent/step spent money.
Trace ID and span ID	Connects Blackridge records to OpenTelemetry traces.
Retry group and attempt metadata	Measures retry amplification.
Fallback group metadata	Measures fallback tax.
Branch contribution flag	Identifies abandoned branch waste.

Minimum Metadata

For useful attribution, send tenant, app, team, workflow ID, workflow run ID, and request ID. For useful lineage, also send root request ID and parent request ID.

Ideal Metadata

For strong forensic reports, add agent ID, agent instance ID, task ID, logical operation ID, trace ID, span ID, retry/fallback group IDs, branch ID, and contributed_to_final_output when known.

What Blackridge Can Reconstruct

With complete metadata, Blackridge can group workflow cost, build request lineage, identify retry groups, estimate fallback tax, find abandoned branches, and tie recommendations to evidence.

Without metadata, Blackridge can still report traffic, provider/model, token/cost data, and some heuristic groupings, but confidence drops.

Related Docs

• Instrumenting Agent Workflows
• Workflow Lineage
• Attribution
• OpenTelemetry Integration