Google has publicly released Deep Review, an AI system built around specialized agents for technical checks of complex documents. Originally focused on computer science research papers, the tool now extends to law and finance. Google says it is designed to help scrutinize content and identify weaknesses while leaving higher-level judgments to people.
From mathematical claims to inconsistencies in legal documents
Google describes a review process that examines logical consistency and mathematical statements, as well as checking factual claims against supporting material. Deep Review is also designed to compare a document with existing work and suggest revisions.
For design documents, Google says the tool can identify unstated architectural assumptions and edge cases and suggest concrete improvements. In presentations, the system is intended to assess whether slides follow a logical sequence and whether conclusions match the charts shown. For legal and financial documents, Google describes checks for consistency between clauses and data tables, ambiguous terms or potential downstream consequences, and claims that need supporting evidence.
Four review stages divide the work among agents
According to Google, processing follows four stages. First, an agent automatically divides the document into logical segments, such as the introduction, methodology and experiments. These segments can overlap, but together they must cover the entire input document.
In the second stage, Google says the system assesses each segment’s information density and allocates a computational budget accordingly. A passage containing demanding mathematical proofs may receive more computing resources than an introduction or a review of related work.
Specialized agents then review the segments in parallel. Google says each agent concentrates on its assigned section while retaining access to the full document, allowing it to follow references and connections across sections.
A final synthesis agent combines the individual reports and checks consistency between sections, according to the company. It repeats logical checks when necessary. This stage also uses external supporting material and tools for verification, which Google describes as an additional measure against hallucinations.
Building on PAT with Gemini 4 Argon support
Deep Review builds on the experimental Paper Assistant Tool (PAT). Google says its pilot programs provided private, pre-submission feedback to authors of more than 16,000 manuscripts who opted in. The programs were conducted in partnership with STOC 2026 and ICML 2026, as well as NeurIPS 2026 and ICLR 2027.
According to Google, this feedback was entirely separate from the official conference review process. PAT did not collect data for training, and papers were removed from Google’s systems shortly after each program ended.
Google is now adding Gemini 4 Argon support to Deep Review. The company says this takes the tool’s capabilities beyond earlier PAT iterations and improves the quality of its feedback and analysis.
API access, input formats and document retention
The service is available worldwide, with every user receiving two free document reviews per week. Users can connect their own Gemini API key for higher limits and programmatic integration through the Deep Review API.
At launch, the tool accepts PDFs and plain text. Google is preparing support for multiple input files, which is intended to enable consistency checks across entire sets of related documents, spreadsheets and exhibits.
Google says uploaded documents are not used to train, fine-tune or improve its models. They are retained for one year solely to display them to the user in the Deep Review interface.
Antigravity gets its own document review pattern
Google also introduced the Teamwork Deep Document Review Pattern in Antigravity, an implementation pattern inspired by Deep Review rather than the service itself. It uses the Teamwork framework and a hierarchy of agents with self-correction and evaluation criteria for individual document segments. Designed to assess research papers, manuscript drafts and technical proposals, the pattern can be invoked by Antigravity users with the /teamwork-preview command.
