ADR 0002: Baseline Fairness

Date: 2026-06-22
Status: Accepted

Context

A major challenge when evaluating multi-agent architectures is proving that the multi-agent decomposition is responsible for a failure, rather than just bad prompting or limited context. In many real-world systems, multi-agent architectures explicitly restrict the context or tools given to any individual agent (e.g., a "CRM Worker" only has CRM tools; a "Ticketing Worker" only has ticketing tools).

If we artificially restrict the baseline single_agent in the same way, we cripple its ability to function as a true control group.

Decision

We will build the single_agent baseline architecture with completely unrestricted access to all tools (CRM, ticketing) and a full, unfiltered context window.

Consequences

By guaranteeing a completely unhandicapped baseline, we establish a clean control. This ensures that when multi-agent architectures fail on the information asymmetry trap (task_02), those failures can be definitively interpreted as architecture-driven (information loss at boundaries) rather than the result of rigging or handicapping the baseline. The single_agent passed 100% of trials cleanly precisely because of this fairness guarantee.