Agent routing should buy verified capability over declared capability
Agent routing — deciding which agent handles a request — operates on declared capability by default: each agent advertises what it can do, and the router picks based on that claim. But advertised capability is cheap to state and even cheaper to misstate. A routing system built from declarations produces the wrong failure pattern — tasks dispatched to agents that look right but act poorly, with no feedback loop to correct the matching.
Capability as a reputation signal, not a registration field
The alternative is to treat capability as earned rather than registered. An agent's routing score should reflect its proven execution history against relevant task classes — not what it claims, but what it has delivered on verified workloads. The multi-agent credit attribution research gives the mechanism: causal tracing links outcomes back through the tool-call chains that produced them, then credits each contributing agent by measurable impact rather than by who happened to be running the pipeline at time of invocation.
In practice, this means:
- Execution reputation over registration: An agent's routing score reflects verified outputs — what it has successfully handled in production workloads, measured by task success and downstream review quality rather than self-reported capabilities.
- Category-weighted scoring: Agents route tasks better when their proven track record is measured within relevant domains (claims processing, not "document tasks") — domain-specific reputation beats aggregate capability statistics.
- Dynamic re-routing based on current performance: A routing pool that shifts weight from agents whose recent score drops below a threshold to higher-scoring alternatives prevents degradation from compounding across cascades.
The credit assignment problem as a routing problem
This is where multi-agent routing gets hard. When multiple agents contribute to an outcome — extraction, classification, summarization — attributing success or failure to any single agent is ambiguous: was the poor output a quality issue from the extractor or a classification misread? Routing decisions that ignore causal attribution (the actual chain value each agent contributed) perpetuate bad matches.
If routing score reflects causal credit instead of declared capability, the system corrects itself: agents that produce good outcomes get more traffic; those whose outputs trigger downstream failures lose share automatically. This is self-improving orchestration — a feedback loop built into the routing logic itself rather than bolted on as post-hoc monitoring.
The practical implication
A reputation-based routing system does not eliminate the design time cost of capability declarations — you still need to know what agents exist and what domains they serve. But it adds an execution-layer check that corrects for misdeclaration after deployment. You declare capacity; you earn distribution. The gap between the two is the space where a routing system either survives or fails.