Your AI counsel, organized the way a legal team is. Sonata arranges its agents into departments spanning deal work, regulatory and compliance, specialized practice areas, and operations, each holding its agents in four clearly marked tiers: Approved agents your department has vetted and tested; agents powered by Sonata that ship free and you copy to make your own; a curated library of Anthropic’s Claude for Legal agents, imported and governed inside Sonata; and My agents, the ones your own people create. A new user reads the trust model at a glance, and access follows roles, so each person sees the departments that are theirs.

General Tools, the built-ins that ship free. Every Sonata workspace comes with a set of general-purpose agents ready to use and yours to copy: a summarizer, a clause and term extractor, an obligations reviewer, a plain-language rewriter, and more. The first one we built, and the one that proves the idea behind the rest, is Document Comparison: give it an original and a revised document and it returns a plain-language explanation of what changed and a visual redline marking every edit in place. Its changes are found by deterministic code, not guessed by a model, so it cannot miss a change or invent one. It is the first of a deterministic core where the part that has to be right is not left to a model’s judgment. The Regulatory department carries its siblings, Bill Version Comparison, the Legislative History Compiler, Hearing Digest, the Docket Analyzer, and the Coalition and Stakeholder Mapper: name a federal bill and the first fetches two published text versions from Congress.gov, lists every version that exists, compares the pair with the same deterministic engine, and explains what changed, never what it means; the second compiles the bill’s record in code, its version trail, full action timeline, and committee reports with their own text, every item cited, and narrates what the record contains without inventing a citation or presenting a committee’s stated purpose as settled intent; the third digests a congressional hearing from its official record, the published transcript when it exists, or the hearing’s known record with an honest note that printed transcripts publish months after the event, every statement attributed as the record attributes it; the fourth condenses a rulemaking docket’s public comments from Regulations.gov in code, grouping exact and near-identical form letters, then identifies the positions and coalitions in what commenters actually wrote, always honest about the comments that arrive as attachments it does not read; and the fifth maps the lobbying landscape around an issue or organization from public disclosure filings on lda.gov, who filed and for whom on what, with amounts exactly as filed, and never a claim about motive, sides, or opposition, because the record does not state them.

Multi-step legal work, built without code and run under human approval. A workflow chains agents into a repeatable process: you assemble it in a no-code builder, start from a template or from scratch, and run it supervised or autonomously. Either way, any step that would change something outside Sonata pauses for a person to approve before it acts. Every run keeps a complete step-by-step record of what ran, what was approved, and by whom. Some workflows run themselves: adopt a watcher, and it runs on the schedule you set, owned by the person who adopted it, under the same approval rules as any other run. The renewal watcher scans your agreements for upcoming expirations; the Federal Register monitor watches for new rules and notices matching the agencies and terms you choose, and can draft a client alert per finding, grounded in the document’s own data, for a lawyer to review, approve, and download; the legislative monitor watches Congress for new bills matching your policy areas and keywords (a plain text match against bill titles and summaries) and follows the bills it surfaces for new actions; and the Comment-cycle monitor watches Regulations.gov for new rulemaking activity matching your agencies and search terms, follows each rulemaking docket that opens a public comment period, and records new documents, deadline changes, and approaching deadlines, always showing the deadline exactly as Regulations.gov states it. The Federal Register and legislative monitors’ findings can become lawyer-reviewed client alerts, and a Comment-cycle finding can become a comment-letter scaffold with the client’s positions left clearly marked for the lawyer, each drafted on demand and grounded in the source’s own record. Nothing is ever sent by the system.