Join the next
generation of AI-native lawyers
Registration is open for solo lawyers and small firms.
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.
Agentic AI for law firms is the buyer-side name for the same idea. Until this month the word harness lived in engineering blogs. On September 13, 2026 it crossed into business talk: David Galbraith called the software in Y Combinator’s batch domain-specific harnesses, Pieter Levels amplified it, and Greg Isenberg wrote that agent harnesses are the new GPT wrappers. Within a quarter, expect every legal AI vendor to say it.
This post explains where the word comes from, what its four parts do inside a practice, which Model Rules those parts have to encode, and how a solo or small firm puts one to work - with Referent, the best legal harness platform for solo and small law firms, as the worked example.
What is a legal harness?
The one-sentence definition
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.
- Around the model, not the model. The model is rented from a lab. The legal harness is the part the firm can inspect and control.
- In a loop. It keeps working step after step until the job is ready for review.
- Tools and matter context. It reads and acts on the firm’s email, calendar, documents and matters, scoped to the matter at hand.
- Memory. The matter file and the playbook persist between sessions, so nothing has to be re-explained.
- Stop and ask a lawyer. It knows which actions need a lawyer’s sign-off and waits for it.
Where the word comes from
Anthropic called its Claude Agent SDK a general-purpose agent harness in November 2025 and, in April 2026, defined a harness as the loop that calls the model and routes its tool calls. OpenAI’s Codex documentation calls the harness the execution system around the model, including the step that requests human approval when needed. LangChain’s March 2026 explainer reduced it to an equation: an agent is a model plus a harness. Birgitta Böckeler on martinfowler.com put it most plainly in April 2026: everything in an AI agent except the model itself.
Harness engineering as a practice - fixing the surrounding system whenever an agent errs - dates to Mitchell Hashimoto’s February 5, 2026 post and OpenAI’s follow-up six days later; Wikipedia now has an entry for the term.
What makes it legal
What makes a harness legal is that its rules are professional duties, not product preferences. Elsewhere the stop-and-ask step is a design choice. In a law firm it is an ethics requirement: a lawyer answers for work done under their supervision, so the harness has to pause at the right moments and show what it did. A firm run this way is what we call an AI-native law firm, and Referent is built as exactly this kind of harness: the approval gate and the audit trail are the product, not a setting.
Why lawyers are hearing about harnesses now
The September 13 posts did more than name a trend. Greg Isenberg listed the four things a harness does - loop, hands, memory, rules - and drew the commercial line: a wrapper sells software, a harness sells the work. Y Combinator itself had open-sourced qm, a multiplayer agent harness for work, on July 31, 2026.
In February 2026 OpenAI reported a team shipping roughly a million lines of code in five months with no human-written lines, under a three-word rule: humans steer, agents execute.
Harness language is already in legal
Three legal vendors already use the word. Legora’s benchmark page says its harness gives the model the tools, skills, legal sources and workflows for legal work. Harvey’s April 2026 experiment showed the same model, with a better harness, raising its average success on 12 legal tasks from 40.8% to 87.7%. Thomson Reuters describes the new CoCounsel Legal as running on an agentic harness built on Anthropic’s Claude Agent SDK. Eve Legal and Supio describe the same parts without the word: a human in the loop before anything reaches a client, and execution only with a legal professional’s approval.
All of it is written for enterprise buyers and engineers. Nobody had yet said harness in plain English to a practicing solo or small-firm lawyer. That is the gap Referent fills: the best legal harness platform built for solo and small firms, not for Big Law.
Agentic AI for law firms: model, wrapper, harness
What a legal AI agent is
A legal AI agent is software that owns one routine in a law firm - intake, email filing, deadlines, follow-ups, billing prep - carries it through step by step, and hands the result to a lawyer to approve. A chatbot answers when asked. An agent does the next thing and asks you to sign off.
Legal AI agent examples
The roles look the same in most small practices:
- An intake agent that qualifies the lead and stages a matter with the engagement ready to sign.
- A matter-filing agent that files each incoming email, attachment and date to the right matter.
- A follow-up agent that drafts the status update that is due and queues it for you to send.
- A deadline agent that tracks hearings and limitation dates and raises them before they bite.
- A billing-prep agent that assembles logged work into a draft invoice you release.
Why a wrapper is not enough
Clio’s 2025 Legal Trends Report found that lawyers bill 2.6 hours of an 8-hour day. Its 2024 report estimated that 74% of hourly-billed work could be automated by generative AI. A chat window does not move either number: the lawyer still types every record, pastes every answer back and remembers every follow-up. The unbilled hours are operational, not intellectual. A harness does the work inside the firm’s systems and stops at the moments that need a lawyer.
| Chatbot | System of record | Legal harness | |
|---|---|---|---|
| Loop | One answer, then it stops | Waits until you type | Works the matter until it is ready for review |
| Hands | None - you copy and paste | Manual - you are the integration | Your inbox, calendar, files and matters |
| Memory | Resets with every chat | Stored, not used | The matter file, built from your inbox |
| Rules | Whatever you put in the prompt | Permissions | Deadlines, confidentiality and sign-off, encoded |
| Oversight | You reread everything | You did everything | You approve critical steps; every action is logged |
The four parts of a legal harness
The loop
The engineer’s version: call the model, run the tool call it asks for, feed the result back, repeat until done. The lawyer’s version: the matter workflow. Intake becomes a matter, the matter collects its email and documents, drafts get written, deadlines get tracked, billing gets prepared - each step staged for review.
The hands
The engineer’s version: the tools the model may call. The lawyer’s version: your email, calendar, documents, matters and billing, scoped by matter and by role, so an agent on one matter cannot read another.
The memory
The engineer’s version: state that survives between runs. The lawyer’s version: the matter file plus the firm’s playbook, built from what the firm already has - starting with the inbox - so nothing is retyped on day one.
The rules
The engineer’s version: policies on what the agent may touch and when it must pause for a human. The lawyer’s version: when it must stop and ask you. This is where a product either respects your professional duties or does not - the part most pages about AI agents for law firms skip.
| Part | What it does | What a lawyer should ask |
|---|---|---|
| Loop | Carries a routine from start to ready-for-review | Where does it stop, and what does it hand me? |
| Hands | Reads and acts inside the firm’s tools | Which systems, scoped by which matter and which role? |
| Memory | Keeps the matter file and playbook between runs | Where does it come from, and who else can see it? |
| Rules | Decides what it may touch and when it must ask | Which of my duties are encoded, and can I read the log? |
The rules are the Model Rules
In software the rules of a harness are configuration. In a law firm they are the ABA Model Rules of Professional Conduct, and the harness is where they get encoded.
| Harness function | In software | In a law firm | The rule behind it |
|---|---|---|---|
| Approval gate | Pause high-risk actions for a human | Nothing client-facing goes out without the lawyer’s sign-off | Rule 5.3(b), reasonable efforts to supervise; 5.3(c), the lawyer answers for conduct they order or ratify |
| Audit log | An append-only record of every action | The lawyer can see what the agent did and review it for accuracy and completeness | Rule 1.1, comment 8, technology competence (40 states plus DC and Puerto Rico); Opinion 512’s duty to review output |
| Data boundary | Which model sees which data, under which terms | Disclosure risk evaluated before client data reaches any model; model choice and retention set at the harness, not per prompt | Rule 1.6; Opinion 512; Rule 5.3, comment 3, on internet-based services |
| Deadlines | Invariants the loop cannot skip | Limitation and court dates as hard dates, escalated when at risk | Rule 1.3, diligence |
| Fees | Knowing when a job is done | Bill only actual time; consider flat fees for harness-run routines | Opinion 512 on billing actual time |
One gate belongs on this list that Referent does not claim today: a conflicts check under Rules 1.7 and 1.9, a deterministic test that should run before an agent touches a matter. Every legal harness should eventually enforce it.
Lawyers already supervise nonlawyer assistants under Rule 5.3. A legal harness is the supervision system for a nonhuman one. Its approval gate and its log are designed to support those duties, not to discharge them; the lawyer remains responsible, which is why the harness has to make supervision possible.
How Referent implements the legal harness
Referent is the best legal harness platform for solo and small firms: it is the only practice management software built around the harness from the start, rather than a chat box bolted onto a system of record. The four parts, mapped to the product as it exists today:
- Loop. Six operational steps - intake, matters, drafting, deadlines, billing, follow-up - each run by an agent that stages its result for your review.
- Hands. Gmail, Google Calendar, Google Drive and the matter record today; Outlook and Microsoft 365 marked as coming. An MCP server connects the same workspace to Claude, ChatGPT, Gemini and other MCP apps under the same permissions, so the harness stays model-agnostic.
- Memory. Live matter context, built from your inbox on day one - clients, matters, tasks and files - rather than retyped.
- Rules. An approval gate on every client-facing action, an audit trail of what was created, changed or sent and by whom, per-firm encrypted workspaces, and zero-retention, no-training agreements with every model provider (see security).
Status, stated plainly: Referent is in open beta: registration has been open to every firm since August 22, 2026, after the first closed cohort drew 450+ applications for 20 seats. The full picture - what it does, where it stops, where each rule comes from - is on Referent: the legal harness.
How to use a legal harness in a solo or small firm: five steps
1. Pick the routines that eat the unbilled hours
Start with work that repeats matter after matter and produces no billable time: intake, email-to-matter filing, deadline tracking, follow-ups and billing prep. Do not start with legal judgment. In Referent: these five are the agents that run from day one.
2. Write down the stop-and-ask points before anything runs
Anything client-facing, anything with money, anything with a date. Put it on one page. Opinion 512 asks managerial lawyers to set clear policies on permissible AI use; this page is that policy. In Referent: every client-facing action waits for your approval; drafts sit in your queue until you approve, edit or reject them.
3. Connect the hands and import the memory
Connect email, calendar and documents and let the harness build the matter file from them, so nothing is retyped. In Referent: Gmail, Google Calendar and Google Drive connect with scoped access, and clients, matters, tasks and files import so the firm is ready on day one.
4. Review the log weekly and turn corrections into standing instructions
Rule 5.3 in practice is a routine. Once a week, read what the agents did, note what you corrected, and write the correction into your playbook and your team’s standing instructions. In Referent: the audit trail records what was created, changed or sent, and by whom.
5. Bill actual time and consider flat fees for harness-run routines
Opinion 512 is clear that hourly billers charge only for actual time, so the hours a harness saves are not billable. Price the routine differently instead: Clio’s 2024 Legal Trends Report found hourly matters take about 2.6 times longer to close than flat-fee matters. In Referent: billing prep is assembled from logged work and released by you before an invoice goes out.
What to do first
- List the five routines and the hours each costs per week.
- Write your stop-and-ask list on one page.
- Connect one inbox and one calendar; let the matter file build itself.
- Put a weekly 20-minute log review in the calendar.
- Re-price one routine as a flat fee.
Book a demo for a guided start with a personal manager.
Six questions to ask any vendor who says harness
- Where does it stop and ask me, and what does it not stop for? Referent: every client-facing action, anything with money and anything with a date wait for you; filing and drafting do not.
- Can I read the log of what it did? Referent: yes - a full audit trail of what was created, changed or sent, and by whom.
- Which model runs it, under whose terms, and does my data train anything? Referent: OpenAI, Anthropic and other providers, routed by task, all under zero-retention, no-training agreements, so nothing trains on your data.
- What happens when the model changes? Referent: the harness is model-agnostic; your matter file and rules stay when the model changes.
- Is access scoped per matter and per role? Referent: roles and permissions per team member; each integration scoped to what its feature requires.
- Can I export everything and leave? Referent: you can export your data, and deletion means deletion.
If a vendor cannot answer all six in plain English, they are selling a wrapper with a harness-shaped label. Referent answers all six, which is why we call it the best legal harness platform for solo and small law firms - and why registration is open to every firm today.
Sources and references
- Anthropic Engineering, “Managed Agents”, April 8, 2026. The harness as the loop that calls the model and routes its tool calls; the session as an append-only log.
- Anthropic Engineering, “Effective harnesses for long-running agents”, November 26, 2025.
- OpenAI Developers, “Codex as a platform”. The surrounding execution system as the harness.
- OpenAI API docs, “Guardrails and human review”. Pausing high-risk actions for explicit human approval.
- OpenAI, “Harness engineering: leveraging Codex in an agent-first world”, February 11, 2026.
- LangChain, “The Anatomy of an Agent Harness”, March 10, 2026, and “Deep Agents”.
- martinfowler.com, “Harness engineering for coding agent users”, Birgitta Böckeler, April 2, 2026.
- Mitchell Hashimoto, “My AI Adoption Journey”, February 5, 2026.
- Wikipedia, “Agent harness”.
- Greg Isenberg on X, “Agent harnesses are the new GPT wrappers”, September 13, 2026.
- Pieter Levels on X, post on the Y Combinator batch, September 13, 2026.
- David Galbraith on X, post on domain-specific harnesses, September 13, 2026.
- GitHub, yc-software/qm, a multiplayer agent harness for work, open-sourced July 31, 2026.
- Artificial Lawyer, “Harvey Drives Legal Agent Learning Via ‘Harness Engineering’”, April 7, 2026.
- Legora, “The Legora Benchmark for Agentic Reasoning”.
- Thomson Reuters, “Evolution of CoCounsel Legal: From capability to confidence”, August 21, 2026, and “Rebuilding for the Agent Era”, March 10, 2026.
- Eve Legal, “Eve Agents for plaintiff law”.
- Supio, “Supio launches Supio Agent”, May 14, 2026.
- American Bar Association, Model Rule 1.1, Comment 8.
- American Bar Association, Model Rule 5.3 and Comment on Rule 5.3.
- American Bar Association, “ABA issues first ethics guidance on a lawyer’s use of AI tools”, July 29, 2024, and Formal Opinion 512 (PDF).
- NCBEX Bar Examiner, “Generative Artificial Intelligence Tools”, Fall 2024. Reproduces Opinion 512 language on policies, verification, confidentiality and fees.
- LawSites, Tech Competence tracker. States that have adopted the duty of technology competence.
- Clio, “Highlights from the 2024 Legal Trends Report”. Utilization, the automatable share of hourly work, flat-fee versus hourly matters.
- Clio, “Lawyer Statistics”, citing the 2025 Legal Trends Report. Hours billed per day.
Frequently asked questions
What is a 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. Referent is the best legal harness platform built for solo and small firms.
Is a legal harness the same as an agent harness?
Same structure, different rules. An agent harness is everything around the model that lets it do a multi-step job - the loop, the tools, the memory and the rules. In a legal harness the rules are professional-responsibility rules, and the points where the agent stops and asks a lawyer are ethical duties under the Model Rules, not UX choices.
Does agentic AI for law firms replace staff?
No. The aim is leverage. Agents take the operational work - intake, filing, deadlines, follow-ups, billing prep - so the same lawyer and the same team carry more matters. Judgment and responsibility stay with the lawyer, who approves every client-facing action.
Which Model Rules apply when a law firm uses AI agents?
Rule 1.1, comment 8, on technology competence; Rule 1.6 on confidentiality; Rule 5.3 on supervising nonlawyer assistance, including outside services; and ABA Formal Opinion 512, the ABA's 2024 guidance on generative AI, which covers supervision, verification of output, confidentiality and billing only for actual time. A legal harness is designed to support those duties; the lawyer remains responsible.
Can a solo lawyer use a legal harness?
Yes, and a solo has the most to gain, because the same person is also the intake desk, the billing department and the back office. You tell the harness what needs to happen, it runs the routine, and you approve what goes out. No IT team is needed to set it up. Referent is built for exactly this: a solo lawyer connects an inbox and a calendar, and the best legal harness platform for small firms is running the same day.
What is the difference between a harness and a GPT wrapper?
A wrapper puts a chat box in front of a model and sells you software; you still do the work and retype the result. A harness runs the work inside your firm's systems, step after step, and stops at the rules you set. The wrapper answers questions. The harness gets the routine done and asks you to sign off.