Auditable Multi-Agent Legal Work

Why it matters

Legal work increasingly runs through pipelines of agents: one retrieves, one drafts, one checks, one plans. Every handoff is a place where responsibility can disappear. A lawyer signing the result warrants the judgments in it, and courts and carriers increasingly ask how work was produced rather than whether it passed a check. If no one can reconstruct which agent made which call and what a human approved, the signature warrants nothing in particular.

State of the art

Single-model tools can log prompts and outputs, but a log is not an audit. Multi-agent frameworks are generally designed for throughput rather than for reconstructing responsibility, and they rarely treat termination as a design problem. Professional guidance is tool-agnostic: counsel must personally verify citations however they were generated, which says nothing about how intermediate machine decisions should be recorded.

Our partial answers

The "tiny company" experiment ran five agents with separated roles (an auditor, a decision-maker, a planner, builders and a test runner), mutual accountability, and red lines that sent any task touching a court-relevant number to a human. It exposed the least understood part of the problem: the loop had no permission to halt and kept running after its authorized work ended, which was fixed by labeling runs productive or idle.

The opposition system design specifies a queue of six attorney judgments, each adopted, rejected or overridden with a logged reason, and gates that warn rather than block. The isResponsive methodology stores every classification with its rubric hash, model version, prompt version and calibration artifact, and refuses to issue a certificate if the audit chain does not verify.

None of this yet adds up to a general standard. Open: a schema for agent roles and handoffs, a definition of which decisions must be reserved to a human, explicit halting criteria, and an audit an outside reviewer can run without the builder's help.