Agentic AI for Legal Research: Building Deep Research in Westlaw and CoCounsel

thomson_reuters · Legal · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Legal research

Thomson Reuters needed legal research agents that plan, adapt, verify, and cite beyond one-shot RAG summaries.

Constraints:

Architecture

Thomson Reuters Deep Research is a legal-research agent inside Westlaw Advantage and CoCounsel. It plans multi-step research, executes searches in parallel, selects tools, opens documents, follows KeyCite breadcrumbs, adapts when findings change, applies stopping criteria, and integrates specialized document-type components for cases, statutes, regulations, and secondary sources. Session memory tracks explored arguments and parallel research threads, while authoritative Westlaw citations, source tabs, excerpts, KeyCite flags, evaluator models, manual review, and component-level model selection form the trust layer.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  lawyer-research-question[Lawyer research question] --> research-planner[Research planner]
  research-planner --> westlaw-tool-layer[Westlaw tool layer]
  westlaw-tool-layer --> specialized-document-agents[Specialized document agents]
  specialized-document-agents --> session-memory[Session memory]
  session-memory --> report-generator[Report generator]
  report-generator --> verification-tools[Verification tools]
  verification-tools --> lawyer-review[Lawyer review]
  evaluation-system[Evaluation system] --> research-planner
  evaluation-system --> report-generator
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
Lawyer research question High-stakes legal query needing research depth and citation support. Research planner
Research planner Generates multi-step plans and alternative theories. Lawyer research question, Westlaw tool layer
Westlaw tool layer Modular search, document opening, KeyCite, and content tools. Research planner, Specialized document agents
Specialized document agents Components tuned for case law, statutes, regulations, and secondary sources. Westlaw tool layer, Session memory
Session memory Tracks explored arguments, related findings, and parallel research threads. Specialized document agents, Report generator
Report generator Synthesizes research threads into a coherent cited report. Session memory, Verification tools
Verification tools Sources tab, inline citations, KeyCite flags, excerpts, and Westlaw navigation. Report generator, Lawyer review
Lawyer review Human verification loop for legal trust and professional judgment. Verification tools
Evaluation system Evaluator models, editorial rubrics, manual review, and model selection. Research planner, Report generator

Data flow

A lawyer submits a legal research question. Deep Research generates a multi-step plan, uses modular Westlaw tools to search and open authorities, routes work to specialized document-type components, tracks findings and arguments in session memory, follows citations and KeyCite warnings, stops when enough support is gathered, and produces a report with inline citations and verification links.

3 supporting quotes
“The system generates multi-step research plans and explores alternative theories when its first approach fails.”article, stated
“Instead of one monolithic agent, we orchestrate specialized components, each tuned for a particular document type (case law, statutes, secondary sources).”article, stated
“All citations in Deep Research come from authoritative Westlaw sources. The Sources tab presents citations visually, with direct links and highlighted excerpts, mirroring how lawyers already work in Westlaw.”article, stated

Why this architecture

The system is agentic because legal research is iterative and authority-driven: lawyers need arguments, jurisdictional variation, statutory changes, and confidence about when enough support has been gathered. A monolithic RAG answer is only a starting point, so Deep Research decomposes work into planning, specialized tools, memory, verification, and evaluation.

3 supporting quotes
“This is not just another RAG (retrieval-augmented generation) tool or search interface, it’s a rethinking of how AI can perform within complex, high-stakes professional workflows.”article, stated
“Legal research is inherently iterative. Skilled researchers plan their approach, reassess findings along the way, and adapt as new information emerges.”article, stated
“the Generation and Verification Loop Despite major advances, LLMs can still make mistakes. Verification is essential in legal research. As Andrej Karpathy has noted, _“the key to AI products is better cooperation”article, stated

Models used: Claude 4 · frontier model providers · evaluator models

Scale

Alternatives considered (roads not taken)

Traditional RAG summary

Rejected because: Lawyers still need deeper authority gathering, edge cases, counterarguments, and jurisdictional analysis.

“A traditional RAG system might suggest the claim can proceed. But that’s not enough. Lawyers want to understand: How strong is the inference?”article, stated

One monolithic legal agent

Rejected because: Document types need specialized components and session-level orchestration.

“Instead of one monolithic agent, we orchestrate specialized components, each tuned for a particular document type (case law, statutes, secondary sources).”article, stated

Patterns

multi-agent-orchestration (legal-research)

Specialized components coordinate across cases, statutes, and secondary sources.

“Instead of one monolithic agent, we orchestrate specialized components, each tuned for a particular document type (case law, statutes, secondary sources).”article, stated

agent-memory (legal-research)

The system maintains research state across long-running workflows.

“The system maintains memory across the session, tracks explored arguments, relates findings, and integrates parallel research threads into a coherent report.”article, stated

safety-policy-enforcement (verification)

Verification tools and citations help lawyers validate generated research.

“Verification is essential in legal research.”article, stated

Tradeoffs

Balance comprehensive legal discovery against result noise.

Chose A bounded agentic research strategy with lawyer verification and context navigation over Unlimited search expansion for every possible authority.

Rationale: Every added search could find a critical precedent or introduce irrelevant material, and legal context windows remained limited.

“The core challenge in legal research remains balancing comprehensive discovery with minimizing noise. Every additional search may uncover a critical precedent or introduce irrelevant results.”article, stated

How much autonomy to grant legal research.

Chose multi-step agentic research with lawyer verification over one-shot RAG answer or unbounded autonomous search.

Rationale: The system needs deeper discovery and adaptation while keeping citations and lawyer review central to trust.

“In Deep Research, the system generates; the lawyer verifies.”article, stated

Gotchas (what bit them)

Traditional RAG summaries were not enough for deeper legal research.

Root cause: Lawyers need strength, counterarguments, and comparisons across authorities.

Resolution: The system plans and adapts multi-step research.

“n New York where the company hired someone of the same race and gender as the rejected applicant. A traditional RAG system might suggest the claim can proceed. But that’s not enough. Lawyers want to understand: How strong is the inference? How does it vary by jurisdiction? What arguments exist on bo”article, stated

Evaluation

The system is evaluated for depth, accuracy, trust, and verification support in legal workflows.

Metrics: depth · accuracy · trust · citation validation

“Together, these three dials give Deep Research real research autonomy and produce results that raise the bar for depth, accuracy, and trust.”article, stated

Outcomes

The summary reports comprehensive analyses take around 10 minutes.

“be/tvpH36uT7hw), an agentic AI system capable of conducting legal research with the sophistication of a practicing attorney, is now live in Westlaw Advantage and CoCounsel.”article, stated