A continual learning benchmark: Meridian Support Agent

A reproducible test-bed for continual learning in a tool-using support agent.

A single source of truth

A fixed company policy and database define every correct action, so ground truth is derived.

Interacting policies

Refund, escalation, entitlement, disclosure, and GDPR rules constrain each other; a local fix can violate a distant one.

Deterministic evaluators

code checks over the final answer and the tool calls

Decisions are tool calls

the agent's real action is which tools it invokes, not just its text

Regression-sensitive by design

tasks are arranged so over-fitting the latest fix is observable as a drop else