Harvey open-sources 10 M&A diligence environments to train legal agents
The synthetic data rooms turn deal review into reinforcement-learning infrastructure, though Harvey has yet to publish agent results.
By Ryan Merket ยท Published
Why it matters
Harvey is turning M&A review into training infrastructure for specialized legal models, extending its business from AI software toward the model and evaluation layer.
Gabe Pereyra (@gabepereyra) and Winston Weinberg (@winstonweinberg) are pushing Harvey deeper into model training, releasing 10 reinforcement-learning environments that ask AI agents to investigate synthetic M&A data rooms and produce cited due-diligence memos.
Harvey announced the release on X on July 17th and published a technical explanation of the benchmark by Julio Pereyra (@ItsJulioPereyra), Harvey's head of legal research. The environments extend Harvey's Legal Agent Benchmark, or LAB, an open-source repository containing tasks, documents, scoring rubrics and an execution harness for testing legal agents.
The release reflects Gabe Pereyra's expanding ambitions for Harvey. Before co-founding Harvey in 2022, he worked as a research scientist at DeepMind and a machine-learning engineer at Meta, according to Harvey's biography. Weinberg supplied the legal half of the pairing after working on securities and antitrust litigation at O'Melveny & Myers. The founders began by experimenting with ChatGPT and legal work. Four years later, Harvey is building the evaluation and training infrastructure needed to produce specialized legal models.
That distinction matters. The new environments do not establish that Harvey can autonomously diligence a live acquisition. They give Harvey and outside researchers controlled places to train agents, inspect their behavior and measure whether they found every planted issue. Harvey said it plans to publish initial results from the M&A agents separately.
A data room built to defeat the context window
M&A diligence forces an agent to do something chat products handle poorly: build a reliable understanding of a business from thousands of scattered records. Lawyers review commercial agreements, employment documents, tax records, litigation materials and intellectual-property files, then connect those findings to deal price, closing conditions and the acquisition agreement.
Harvey's environments recreate that process with synthetic virtual data rooms. Agents receive deal context and a filesystem, investigate the records, identify legal and commercial risks, and draft a diligence memo. The output is graded against rubrics containing known findings and recommendations planted in the environment.
Harvey says its largest released environment contains 80 million tokens of context and 1,000 grading criteria. That is different from the 1,000-document figure in the supplied Aligned News summary. Harvey's detailed post describes another example, a fictional acquisition of Sentinel Cloud Security by Helios Cloud Holdings, containing more than 3,500 documents and about 45 million tokens. Harvey has not published a document-by-document accounting that reconciles the two examples, so the figures should be treated as descriptions of separate environments.
The scale is central to the exercise. An agent cannot place tens of millions of tokens into an ordinary context window and reason across them in one pass. It must decide what to read, preserve facts over a long run, revisit evidence and connect clauses found in different files. It also has to resist the retrieval habits rewarded by many existing agent tasks. A diligence lawyer cannot stop after finding several relevant documents. A missed change-of-control clause, consent requirement or ownership gap may alter the economics of the transaction.
Harvey's public LAB tutorial shows the smaller end of that structure. Its sample M&A assignment gives an agent 60 synthetic documents and grades the resulting red-flag memorandum against 68 criteria. The repository is licensed under MIT and supports models from Anthropic, OpenAI and Google through its harness.
The benchmark is part of Harvey's model-layer bet
Harvey initially released LAB on May 6th with more than 1,200 tasks across 24 legal practice areas and over 75,000 expert-written rubric criteria. The M&A environments move the project toward longer jobs that resemble complete client matters rather than isolated contract questions.
That progression supports Pereyra's stated model strategy. He has said Harvey is working on a legal foundation-model series designed for matters that can run for months and involve dozens of associates. In that plan, an agent coordinates legal tools and sub-agents, escalating to frontier models or human partners when needed. Open benchmarks give Harvey an external test of whether post-training improves that system while supplying structured environments for reinforcement learning.
Harvey also has a commercial incentive to define the benchmark. A widely used evaluation can steer model developers toward the workflows Harvey already sells to law firms and corporate legal departments. It can also generate research around Harvey's preferred unit of work: the complete legal deliverable, graded as a lawyer would review it.
The benchmark remains Harvey-designed. Harvey selects the tasks, generates the synthetic documents and defines the scoring rubrics. That produces clean reward signals for training, while leaving open how performance transfers to confidential transactions with incomplete disclosure, inconsistent file organization, shifting document sets and disagreements among experienced lawyers.
Harvey's own initial LAB results, published May 26th, showed frontier models completing fewer than 10% of tasks end-to-end under its all-pass standard. A task fails if any required criterion fails. The strict method matches the consequences of missing a material issue, though it also means the reported scores are measurements under Harvey's rules rather than an independent industry evaluation.
Capital for a move beyond legal software
Harvey's research push follows a March 25th financing that brought in $200 million at an $11 billion valuation. GIC and Sequoia co-led the round, with Andreessen Horowitz, Coatue, Conviction Partners, Elad Gil, Evantic and Kleiner Perkins participating.
The funding announcement said Harvey would use the capital to expand the agents customers run on its platform and grow its legal-engineering operation. LAB shows another destination for that money: synthetic data generation, evaluation systems, model post-training and the compute required to run agents across data rooms containing tens of millions of tokens.
For Weinberg, the former litigator, the product thesis has been that lawyers should retain responsibility for judgment and strategy while agents execute longer workflows. The M&A environments put a measurable target behind that pitch. Harvey is defining what an AI associate must read, produce and verify before a law firm can trust it with a transaction.
The hard evidence will come from performance, including whether trained agents can exhaustively review these environments at an acceptable cost and whether improvements survive contact with live matters. For now, Harvey has released the test and the training ground. It has not released a passing diligence agent.