GenieAI Launches Agents That Build Their Own Teams and Tools for Each Legal Task

GenieAI has launched Legal Specialist Agents, a system in which a lead agent assembles a bespoke team of AI specialists for each legal request instead of running it through a fixed workflow. The London and New York company says the agents also write their own tools at runtime, a design choice that separates it from most contract-focused legal AI products.

Fixed workflows are the weak point

Most legal AI tools ask a customer to design a workflow first, then push every request through it. That works for repeatable tasks such as standard NDAs. It strains when a matter brings unusual parties, paper and risks. GenieAI is betting that planning at the level of the individual request, not the product, is the next step for legal technology. The pitch also targets a complaint legal teams voice often: output that varies from one day to the next.

The agents scale the team to the task

GenieAI says a two-page NDA gets two agents and one quality gate, while a 40-document due diligence project draws 12 agents and three review gates. Each agent holds a legal role, such as drafting, negotiation, risk analysis, research or proofreading. Specialists write short programs to read documents and tie findings to sources. Each tool is checked for coverage, reviewed by a second agent and run in a read-only sandbox. Agents also consult approved legal websites for current law in the relevant jurisdiction, and exceptions go to a named person.

The release also fixes a practical irritant. Genie now edits the original Word file with formatting intact, keeps every version and lists open issues by severity. Customers can send documents for signature through DocuSign and Adobe Sign, and the product connects to Gmail, Outlook, Google Drive, SharePoint and OneDrive.

GenieAI cites two head-to-head reports against Claude Cowork. In a contract renegotiation, its agents passed 11 of 11 checks against eight for the rival. In a 57-document force majeure dispute, they passed 15 of 15 against nine. These are vendor-published results, though the company says every check is public and independent reruns are supported.

Benchmarks now shape buying decisions

The comparison with Claude Cowork is notable. It positions a specialist vendor against a general-purpose assistant from a frontier model developer, a matchup that enterprise legal buyers increasingly raise in procurement. Specialists must show that domain design beats raw model capability. Published, rerunnable test suites are one way to make that case, and rivals may feel pressure to disclose their own methods.

The architecture claim carries weight beyond marketing. If agents can plan, build tools and verify one another, the cost of adapting legal AI to a new matter type falls. That would favour vendors with strong orchestration over those with large libraries of preset workflows. GenieAI also says a small engineering team built the underlying technology with 75 agents, a claim that points to a leaner cost structure for vendors that adopt similar methods.