Multi-agent economics fail without causal credit assignment
Pricing agent work starts from a simple assumption: each task has a cost, and each agent contributes some amount to the total. In single-agent systems this is trivial — one model call, one billable unit. In multi-agent pipelines you can have five or more agents touching a single user request, chaining tool calls across subsystems, and the attribution problem becomes material.
The attribution chain
Consider an end-to-end claims pipeline: Agent A extracts text from a PDF (100 tokens), Agent B classifies it into categories (50 tokens), Agent C scores it against rules requiring 10 function calls, and Agent D generates the final decision letter. The user pays for all five agents' runs. Who should be billed to whom? Which agent delivered value versus consumed cost?
Causal credit assignment traces each successful outcome back through the tool-call graph that produced it — not just forward from inputs or backward from outputs, but through every causal link in between. In formal terms: credit flows only along edges where an action was both necessary and sufficient for the observed outcome. This eliminates two common distortions:
- Affiliation billing: Charging the customer for every agent that ran, regardless of whether its output materially changed the result.
- Fair-share averaging: Splitting cost evenly across contributing agents when one actually supplied 90% of the work and another consumed a few tokens on an irrelevant side branch.
The economic implication for agent fleets
Economics without attribution is price fiction. Without causal credit, you cannot tell which agents are profitable — only which ones are running. You cannot optimize cost because you cannot distinguish the expensive path from the wasteful one. You cannot negotiate with clients because your per-task cost collapses into an aggregate total that contains no leverage points.
Auditability compounds the problem. An agent that produces a bad outcome but carries causal credit for a good chain is both over-rewarded and under-exposed: rewarded through inflated attribution, exposed through failure analysis. The reverse happens to agents that do nothing materially but carry execution tokens: they get credited for nothing.
A practical approach
The minimum viable credit assignment traces each output back to the tool-calls graph, then scores every node on how much it changed the outcome — a counterfactual measure ("what would this task's result have been without this agent?"). Agents scoring above threshold get full cost attribution; those below do not. In high-stakes environments, only agents that contributed causally can recover their own costs.
It is not free: attribution tracing adds latency to every billing cycle. But an unattributed fleet produces the wrong optimization signal — either under-investing in the right tools or over-investing in the wrong ones. Attribution turns pricing from a cost recovery exercise into an architectural decision.