AcceptedADR-001
Thompson Sampling
Context
Routing agents without historical data leads to suboptimal allocation. A probabilistic approach was needed.
Decision
Each (phaseType, agent) pair maintains Beta(α,β). We sample θ and pick the highest. α increments on success, β on failure.
Consequences
Routing improves over time. Cold start is randomised. Requires persistent storage (ADR-003).