DeterministicCoordinationOverAgentAdjudication
title: Deterministic Coordination over Agent Adjudication radar_quadrant: Techniques radar_ring: Assess
Multi-agent AI pipelines commonly use an LLM not just to reason about a task, but to decide what should happen next: which agent runs, whether a result passes, how work gets routed. That coordination role feels natural to hand to an agent, but it is usually pure bookkeeping rather than judgment, and every LLM call used for it is paid for at the same token rate as genuine reasoning. Adam Jacob's guide to reducing token spend describes a technique for separating the two: extract the coordination logic — routing, adjudication, state tracking — into ordinary deterministic code, and reserve agent calls for the smaller set of decisions that actually require evaluation or reasoning. As the post frames it, "you use the Agent to build the program that minimizes the need for the Agent itself."
The technique is demonstrated on a code-review pipeline nicknamed Garfield. In its original form, the pipeline used 23 separate agents to coordinate a review, at a cost of roughly 4.6 million tokens per run. After moving the routing and adjudication logic into typed, code-based workflows and limiting agent invocations to the steps genuinely requiring judgment, the same pipeline ran with 3 agents and roughly 500,000 tokens — an 8x reduction in token spend, in about half the wall-clock time, with no loss of functionality.
This is placed in Assess because it is a documented, demonstrated technique with a concrete before/after result, but the ring requires the user's own confirmed production use to move further, and no such use exists yet. The underlying principle — treat agent coordination as a code-design problem, not a prompting problem — generalizes beyond the specific Garfield example and is worth trialing on any multi-agent workflow that has grown expensive.
