The problem
A router should produce evidence, not just an agent name
In many agentic systems, a request arrives, candidate agents or tools receive scores, and the highest-scoring option wins. That is enough to execute a workflow, but it is not enough to explain or govern one.
When delegation fails, an operator needs to reconstruct the evidence considered, verify that the selected candidate was permitted and available, understand why alternatives were rejected, and determine whether the system should have clarified, abstained, or escalated instead.
Evidence is not eligibility. A score is not a probability. Probability is not utility. Policy is not intent. Explanation is not provenance.
The proposal
A decision-provenance lifecycle for delegation
Evidence-Grounded Agent Routing treats routing as a small decision system. Hard constraints are evaluated before optimization, confidence is calibrated, and the router can choose a safe non-selection action when the available evidence is insufficient.
- 01RequestIntent and constraints
- 02CandidatesRecall before ranking
- 03EvidenceSource-linked snapshots
- 04EligibilityPolicy and operations
- 05DecisionSelect or defer
- 06OutcomeEvaluate and recalibrate
The resulting action can be select, parallel, clarify, abstain, or escalate. The architecture avoids treating forced selection as the only successful outcome.
The artifact
The Routing Decision Record
Every consequential route produces a versioned Routing Decision Record, or RDR. It connects the pre-decision evidence to the eventual outcome and preserves what is needed to replay or challenge the delegation.
Executable evidence
What the reference implementation demonstrates
The companion repository implements the control path in Python and LangGraph with typed records, policy and operational gates, configurable scoring, abstention, provenance capture, replay checks, and automated tests.
These are development results from 42 authored synthetic cases: 12 calibration cases and 30 evaluation cases. They demonstrate that the architecture is executable and that its safeguards affect behavior within the tested scenarios. They are not independent evidence, production validation, or a claim of general superiority. One known evaluation case remains visible in the repository.
Meaningful uniqueness
The contribution is integration, not invention by renaming
Capability matching, policy engines, confidence calibration, abstention, provenance, and evaluation each have established histories. The paper’s contribution is to integrate them into one operational lifecycle for agent delegation.
The router no longer emits only an agent identifier. It produces a falsifiable decision claim: given this request, candidate set, evidence, policy, and system version, this action was eligible and preferable—and the record exists to examine that conclusion later.
Open artifacts