Multi-Agent Observability: When Your Agents Orchestrate Other Agents

Multi-agent systems where agents orchestrate other agents create cost attribution challenges that single-agent observability cannot solve. Here is what changes.

How Multi-Agent Cost Propagates

In a multi-agent system, an orchestrator agent receives a task, breaks it into subtasks, and delegates those subtasks to specialized subagents. Each subagent makes its own model calls. The total cost of the orchestrated task is the sum of the orchestrator's calls plus every subagent's calls — but that total is almost never visible as a single unit in standard monitoring.

The Attribution Challenge

When an orchestrator invokes a subagent, the subagent's model calls typically appear as separate sessions in your observability data. Without explicit parent-session linking, you cannot reconstruct the full cost of the orchestrated task. You see the pieces but not the whole. A task that costs $4.20 when you add up all the subagent sessions may appear as four separate $1.05 sessions with no obvious connection between them.

What Proper Multi-Agent Instrumentation Requires

Multi-agent observability requires a parent-child session model. Every subagent invocation should carry a reference to the parent session that initiated it. This allows the observability layer to aggregate all child session costs into a single task-level cost record, regardless of how many agents contributed to the result. Without this, your cost data is always incomplete.

Loop Amplification in Multi-Agent Systems

Agentic loops become significantly more expensive in multi-agent systems because a loop at the orchestrator level triggers multiple subagent invocations per iteration. An orchestrator that loops 10 times while delegating to 3 subagents per iteration generates 30 subagent sessions, each with their own model calls. The cost amplification is multiplicative, not additive.

Practical Starting Points

Most teams building multi-agent systems do not instrument for multi-agent observability from the start. The practical approach is to add parent-session ID propagation to your agent invocation layer, establish total task cost as a first-class metric alongside individual session costs, and build alerting on task-level cost rather than session-level cost for orchestrated workflows.