Why an AI Legal System Needs More Than One Agent

The Problem

A general-purpose AI agent (e.g., a chatbot) is useful for answering questions. Legal work, however, rarely consists of one question followed by one answer.

Typically, a substantive legal task may require documents to be summarized, specialized analysis to be performed, intermediate results to be reviewed, feedback to be incorporated, and a final work product to be reviewed by an attorney.

Asking one general-purpose AI agent to perform all of this in a single interaction makes the process difficult to track, review, revise, and evaluate.

The Approach

Here, we are building legal AI systems around workflows rather than isolated AI agents.

Consider an invention disclosure analysis during patent application filing. Instead of asking one general-purpose AI agent to simply “analyze this invention,” our system decomposes the work into multiple stages. For example, a pipeline managing agent coordinates the overall execution and distributes tasks to specialized agents for disclosure intake, prior-art research, invention analysis, compliance review, feedback analysis, feedback review, and report generation. Each agent operates within a defined stage and produces work products that can be passed to subsequent stages, reviewed, or revised.

The important distinction is not simply the number of agents. More agents do not automatically make an AI legal system better. The value comes from giving each stage an explicit responsibility and connecting those stages through a structured process.

The outputs of those stages also become work products that can be versioned, reviewed, and passed to a subsequent stage for further review and revision.

What We're Investigating

Our current question is not whether a collection of specialized agents can produce an impressive legal work product. It is whether a structured workflow can make AI-assisted legal work more explicit, reviewable, and testable.

That means we are investigating questions such as:

• Does decomposing a task improve the quality of the resulting work?
• Which tasks benefit from specialization?
• What information should move between workflow stages?
• Where should attorney feedback enter the process?
• Which intermediate work products are actually useful to an attorney?
• Where do errors propagate from one stage to another?

We expect some of these design choices to fail. That is part of the investigation.

Why It Matters

Attorney review is not a disclaimer added after an AI answer is generated. In our system, it is a workflow stage with a defined purpose: reviewing AI findings, adding comments and legal analysis, and providing feedback that can feed back into the process. In other words, attorney review is built into the workflow.

A structured workflow also gives us something that a single opaque answer does not: places to evaluate the legal AI system. We can examine individual stages, intermediate work products, revisions, and the final work product.

Conclusion

This architecture does not make legal AI reliable by itself. Multiple AI agents can introduce new failure modes, including coordination errors, inconsistent outputs, and propagation of mistakes between stages.

The reason to build a structured legal AI system is not to assume those problems disappear. It is to make the work visible enough that we can observe, test, revise, and learn from it.

That is the hypothesis we are investigating: useful legal AI may depend less on finding a single agent that can “do legal work” and more on engineering a system in which AI work can be decomposed, preserved, reviewed, revised, and evaluated.