Tom Smoker uses graphs to make legal-agent workflows explicit, testable, and editable as information changes.
2
WhyHow combines web scraping, machine learning filters, large language models, and lawyer-specific schemas to find potential cases early.
3
Human review and graph-based state management help chained agents reach a level of accuracy that lawyers can use.
Summary
Tom Smoker explains how WhyHow uses graphs and multi-agent systems in legal work. Legal applications need exact language, traceable decisions, and the ability to locate where a workflow went wrong. A sequence of probabilistic agents can lose accuracy at every step, so the system needs structured state, guardrails, and human review. WhyHow scrapes online complaints, filters them through schemas set by individual law firms, and generates reports in the format each lawyer expects. Graphs connect people, products, ingredients, dates, complaints, and legal claims, allowing lawyers to query and inspect the relevant subgraph. Smoker also describes discovery workflows, where structured extraction can reduce a huge document collection to information an expert can assess. He is candid that this is not an LLM-only system. Traditional machine learning filters the data, while LLMs help connect parts of the workflow that were previously too inconsistent to automate end to end.
Graphs make relationships explicit enough for legal analysis
Smoker defines graphs through their relations. They let a user see what is connected to something else, state those connections explicitly, and run large-scale analysis over them. The links can span several steps, from node to node to node, and can branch in different directions. In a mass-tort example, a graph can connect individuals to products, ingredients, concentrations, identification numbers, dates, and harms. That structure gives legal experts a visual and queryable way to focus on the parts of a case that matter.
A multi-agent system is a workflow with controlled state
Smoker uses multi-agent systems to break a complicated white-collar workflow into specific steps. Each step can have different inputs and outputs, requirements, frequencies, and state. In WhyHow's systems, a graph can control that state. He describes a cycle of capturing information, expanding it, pruning it, structuring it, and querying it. This approach gives the team a way to inspect and modify the information held across the workflow instead of relying on an LLM's immediate context alone.
Legal agents need more control than an LLM can provide by itself
Lawyers expect information to be correct and written in the proper language. Smoker says probabilistic language models do not meet that requirement in isolation. The system must have a specific structure and schema, and the team must be able to test it and identify what is going right or wrong. He is also skeptical of the word agent, since a bad prompt can quickly produce a bad agent. WhyHow spends time on guardrails, episodic memory, state capture, and state pruning.
Sequential accuracy falls as agent steps accumulate
Smoker gives a simple calculation for chained workflows. If one agent is 95 percent accurate, he considers that a difficult target already. If five agents with that accuracy run sequentially, the expected accuracy at the end is 77 percent. He uses this to describe the central engineering problem as decision-making under uncertainty throughout the workflow. Guardrails and humans in the loop can bring the result to a level that people are willing to use, even when individual agents are not perfectly reliable.
Structured discovery turns document volume into inspectable evidence
Legal discovery can deliver hundreds of gigabytes of emails, much of which may not matter to a case. Manual review is slow, and lawsuits often run on fixed timelines. Smoker describes extracting and structuring the documents so that irrelevant material can be dismissed and useful information can be inspected. The graph can augment the extracted information and present it visually to an expert. In a pharmaceutical example, the system could connect an ingredient and its concentration to a problem, a time, and the small number of people affected.
Case research depends on each law firm's definition of a useful signal
WhyHow scrapes online sources where people discuss possible harms, including government websites, forums, subreddits, and other communities. The company then filters the material through schemas created with particular lawyers and firms. Smoker says there is no universally perfect case because the right opportunity depends on the partner, firm, jurisdiction, claim, and other preferences. A lawyer-specific graph and agent workflow can preserve those preferences, then generate a report for that lawyer.
Complaint density can reveal vehicle cases before they become obvious
Smoker describes tracking reports of vehicles that smoke or catch fire. The system structures complaints by vehicle, model year, and problem, then calculates density from the number of complaints and their velocity over time. Previous public lawsuits provide additional information about how similar cases developed. He says a large firm might take a case eight or nine months after complaints begin, while WhyHow can find a potential signal in about 15 minutes and may take roughly a month to gain confidence in it.
Reports are often more useful to lawyers than chat answers
Smoker says WhyHow uses less of a chat-based RAG interface. Lawyers may want an answer, but they also want the form of work product they already use. Each day, the system builds and updates graphs, then composes a relevant subgraph into an actionable report. The graph schema remains extensible, so lawyers can update filters, provide feedback, and query the information across the broader structure. The output improves as the system manages and tracks each lawyer's preferences.
LLMs connect an ML-filtered system rather than replace traditional methods
Smoker rejects the idea that generative AI is automatically better than machine learning. Traditional multi-agent systems often stopped when two workflow components were too probabilistic or inconsistent to connect. LLMs make those connections possible, including cases where the desired input is expressed in natural language rather than as a binary feature. WhyHow still uses machine learning to filter information and uses LLMs to pipe parts together. Graphs provide the state that can be iteratively built and pruned.
"It is an ML filtered system that LLMs have allowed us to pipe together such that you can actually provide value completely end to end."18:13
Who should watch
You are building legal or other high-accuracy agents and need to inspect, test, and edit state across a multi-step workflow.
Your system processes large document or web collections, but domain experts need structured evidence and reports instead of raw search results.
You want a production example of combining graphs, traditional machine learning, LLMs, and human review rather than treating an LLM as the whole system.
