Referent: the legal harness

A legal harness is the software layer around an AI model that turns it into a working agent for a law firm: it runs the model in a loop, gives it the firm's tools and matter context, keeps the matter file and the firm's playbook as memory, and decides when the agent must stop and ask a lawyer.

The loop is the matter workflow: the agent works step by step - intake, engagement, first drafts, follow-ups, billing prep - instead of stopping at a chat reply. The hands are the tools your firm already runs: your inbox, calendar and files, plus matters, documents and deadlines inside Referent.

The memory is the matter file, built from your inbox, so every email, document and deadline lands where it belongs. The rules are the Model Rules, encoded: nothing client-facing goes out without your approval, deadlines are hard dates, and every action is logged.

A chatbot answers when asked. You paste the context in, copy the answer out, and update your own systems by hand. When the chat ends, the work stops and the memory resets.

A legal harness owns a routine and keeps working: it reads the inbox, files each email to its matter, drafts the reply, tracks the deadline and prepares the bill, then stages each result for your review. It works from your live matter file rather than a blank prompt, so its output is about your actual clients and dates. And it stops at the rules: nothing client-facing leaves the firm until you approve it, and every step is logged.

That is the difference between a wrapper that puts a chat box in front of a model and a harness that puts the model to work inside your firm. Chatbots answer questions. A legal harness runs the operation, which is what makes a firm AI-native rather than AI-assisted.

Nothing in the Model Rules forbids a lawyer from using AI agents, but several rules govern how. Rule 1.1, Comment 8, asks lawyers to keep abreast of the benefits and risks of relevant technology, a duty 40 states plus DC and Puerto Rico have adopted. Rule 5.3(b) and (c) make a lawyer responsible for supervising nonlawyer assistance and for conduct the lawyer orders or ratifies.

ABA Formal Opinion 512, issued in July 2024, applies those duties to generative AI: review the output rather than rely on it, protect client confidentiality, set firm policies on permissible use, and bill only for actual time.

Referent's approval gate and audit log are designed to support those duties: every client-facing action waits for your sign-off, and every step is recorded so you can review it. The lawyer remains responsible for the work. The harness makes the supervision practical instead of leaving it to memory.

Referent stops at three kinds of moments: anything a client will see, anything that touches money, and any deadline at risk.

A client email needs a reply - the draft is ready, and nothing is sent until you approve it. A new lead finishes intake - the matter is staged with the engagement ready to sign. A first draft is done - it waits in the matter for your edits and sign-off. A deadline is close - it is tracked as a hard date and surfaced to you, not buried in a thread. An invoice is ready - billing is prepared in the background, and you release it. A follow-up is due - it is written and queued, and you send it.

Between those moments the agents keep working: reading, filing, drafting and tracking. Every action, including the ones that did not need your approval, is logged in an audit trail you can inspect at any time.

No. Referent does not train AI models on your firm's data, and it does not sell, share or monetize client information in any form. Your firm's data lives in an encrypted workspace that is isolated per firm, so one firm's information is never mixed with another's. Where the agents call a model vendor, Referent's agreements with those vendors are zero-retention and no-training: the request is processed and not kept.

Inside the workspace, access is controlled by role, so each team member sees only what their role allows, and a full audit trail records what was created, changed or sent, and by whom. When you delete data, it is deleted, not archived out of sight.

Confidentiality is the design constraint the harness is built around, because for a law firm the client's trust is the asset that matters most. The full picture, including encryption and certifications, is on the security page.

The harness is model-agnostic. Referent's agents run on models from more than one vendor - the security page lists the zero-retention agreements with OpenAI and Anthropic - and the model behind an agent can change without changing anything you see. Your matter file, your rules, your approval gate and your audit trail live in the harness, not in the model. That is the practical point of separating the two: when a better model ships, the harness keeps working and the firm keeps its memory.

It also works the other way. Referent runs an MCP server, so the AI apps you already use - ChatGPT, Claude, Gemini, Perplexity and any other MCP-compatible app - can read and act on your matters through the same permissions, approvals and audit trail as the Referent app.

The model you talk to can be your choice; the rules it works under are the firm's. Setup for each app is on the MCP page.

Referent is built for solo lawyers and small firms - the practices where the lawyer is also the intake desk, the billing department and the back office, and where a harness that runs those routines changes the math immediately.

Registration has been open to every firm since August 22, 2026. Before that, the private beta started as a cohort of 20 seats, and more than 450 firms applied for them in under four weeks; that demand is why we opened up.

A one-lawyer practice gets full firm leverage: intake, email, deadlines and billing prep run from the matter file while the lawyer approves what counts. A firm of two to ten lawyers carries more matters with the same team. Big Law support is coming later. If you want to see how the harness fits your size of practice, start with the solo law firm page or the small law firm page.