When one AI agent's work becomes another agent's input, who should own the cost?

I've been thinking about a problem that gets interesting once agent systems stop behaving like isolated runs.

Imagine a research agent spends $0.40 gathering information and produces an artifact. Five other agents reuse that artifact instead of doing the research themselves.

The accounting seems straightforward until you ask what you're trying to measure.

If the research agent gets charged the full $0.40, its cost looks high even though its work saved the system from repeating it.

If the five consumers each get charged the full amount, total reported spend becomes misleading.

If you split the cost equally, you might get a neat dashboard that tells you very little about which architecture is actually better.

And if the artifact is reused a hundred times, the economics change again.

I can see at least three different numbers worth tracking:

• Execution cost: What did this agent actually spend?

• Marginal cost: What additional expense did this particular use create?

• Allocated cost: How much of the shared work should count toward this consumer or workflow?

These numbers answer different questions. The first helps with debugging, the second with optimization, and the third with budgeting and accountability.

But I'm not convinced there's one allocation rule that works for every system.

For those building multi-agent workflows, shared context, retrieval pipelines, or reusable agent outputs:

Would you keep these as separate accounting views, or establish one canonical cost attribution model?

And when comparing two architectures, which number would you trust to tell you that one is genuinely more efficient?