Multi-Agent Legal Workflow

Question

Does coordinating multiple specialized AI agents produce a better-organized legal work product than a single general-purpose AI assistant — and what coordination overhead does that introduce?

Objective

Test whether decomposing a legal workflow across specialized AI roles, rather than one general-purpose assistant, improves task organization enough to justify the added coordination complexity.

Background

The Workflows page's AI Team section describes organizing AI participation in legal workflows around specialized roles rather than a single assistant. This experiment is where that idea is meant to be tested directly, rather than only described.

Hypothesis

Specialization should improve the legibility of intermediate work products, but multi-agent coordination is expected to introduce its own failure modes — the experiment is designed to surface which effect dominates in practice.

System

Conceptually: a workflow manager assigns parts of a matter to specialized AI agents, whose intermediate outputs are then reviewed and revised. See the Workflows page's AI Team section for the current conceptual architecture this experiment is designed around.

Method

TBD — the experiment's method will be documented once it has been run.

Results

TBD — no results have been published yet.

Evaluation

TBD — evaluation criteria will be published alongside results.

Failure Modes

TBD — failure modes will be documented as they are observed, including coordination failures specific to multi-agent workflows.

Human Review

TBD.

Lessons

TBD — lessons will be published once there is something to report.

Next Experiment

TBD.