The Org Chart Is Now an Architecture Diagram.

On 7 July 2026, Norm Ai announced a US$120 million Series C at a US$1.2 billion valuation, led by Khosla Ventures.¹ That made it the category's first unicorn, though the valuation attaches to the platform company. The law practice is its affiliate, Norm Law LLP, launched in November 2025 alongside a US$50 million Blackstone investment; it describes itself as senior attorneys supervising AI agents rather than managing associate pyramids, with pricing tied to outcomes rather than hours.²

Three days later, six researchers in Japan and Vietnam posted a paper called L-MAD: by its authors' positioning, the first systematic evaluation of multi-agent debate structures on a legal reasoning task.³

The funding says the category is real. The paper says, for the first time with legal-task evidence, what running one well requires.

Lawyers supervise agents. But supervise how? Any system of agents has an escalation path and a knowledge boundary, whether or not anyone chose them. Build it as a panel that argues and votes — the shape the AI-native firms describe — and it acquires a roster, a round budget and a vote protocol as well. In a firm made of agents, the org chart is an architecture diagram and the supervision policy is a config file. Someone has already set those values, at every firm in this category. Nobody has published them.

One caveat before anything else. As at 25 July 2026, no AI-native firm has disclosed revenue. The honest evidence for this category is capital flows and structural design, not outcomes.

Artificial Lawyer calls these firms NewMods: legal services businesses designed around an AI operating layer from day one, senior lawyers supervising and calibrating agents instead of running the traditional leverage model, pricing decoupled from hours. The category barely existed before 2025. By mid-2026, Artificial Lawyer was tracking more than 27 such firms.⁴

Three of them tell the story.

The incumbents can see the shape changing. Citi's 2026 client advisory put it in one line: the shape of leverage is likely to change “from a pyramid to a cylinder” — fewer juniors, more senior lawyers, as generative AI absorbs the repetitive base. In the same survey, 63% of large firms said they expect generative AI to affect their leverage models within a decade, up from 43% a year earlier.⁷

The NewMods propose a third shape. Not the pyramid thinned into a cylinder, but a lawyer above a panel of agents, with a layer of rules in between.

Almost all of that is structure and money. Garfield's trial win is the category's one public outcome to date, and one small-debt trial is not a track record. So look instead at what the evidence says about the structure itself.

L-MAD comes from six researchers at JAIST and VNU University of Engineering and Technology in Hanoi.³ The setup: three generated expert personas debate whether a legal hypothesis follows from articles of the Japanese Civil Code. Binary answer, entails or does not entail, on an established benchmark. The researchers tested two families of protocol — forced consensus, where agents argue until they agree, and voting, where agents reason independently and then vote.

Read the caveats first. The models tested are small open-weights models, nowhere near the frontier systems commercial legal products run on. Binary statutory entailment is about as narrow as legal reasoning gets. It is a version-one preprint from a single group reporting its own results, though the arXiv listing notes it as an outstanding paper at ICML 2026's AI4Law workshop. And the strongest model tested gained nothing from debate, because it had already saturated the task alone — so the firms running frontier models probably have the least to gain from debating. The structure transfers. The numbers do not.

Read together, the four findings are not four facts about small models. They are four decisions, and every agent-run firm has already made them.

Who may debate is a hiring decision. Because of the competence floor, the model behind each agent has to be chosen for capability first and price second. Staff the panel below the floor and the debate produces confident, wrong answers.

How long they argue is matter management. Once the panel is unanimous, more rounds only give agents room to argue themselves out of correct answers. A round budget and a stop rule are the agent version of knowing when work is finished.

How they vote determines whether disagreement survives. Blind, independent votes preserve the information in a split. A protocol that pushes agents toward each other destroys it.

When a human gets the file is the escalation policy, and the paper hands firms a mechanical trigger: any split. Routing splits to a lawyer is a defensible middle ground between reviewing everything and reviewing nothing.

What the panel cannot know is the knowledge boundary. It is the setting most likely to be neglected, because it lives outside the agent system, in the firm's knowledge layer.

For Australian readers, none of this is a future governance frontier. Solicitors carry competence and supervision duties under the Australian Solicitors' Conduct Rules, and a joint statement by three Australian legal regulators in December 2024 made verifying AI output an expectation under the existing competence duty. No new rule was required.⁸

What changes in a firm of agents is the form the duty takes. The floor, the budget, the protocol and the escalation rule become values in a system, and values in a system can be versioned, logged and audited. The literature on agent infrastructure has a name for where this ends up: executable accountability — approvals and escalations recorded as durable system state rather than habits.⁹

The labour market and the insurance market are drifting to the same place. James Tuke argues that junior training in AI-heavy firms becomes an “apprenticeship of oversight”, where juniors learn to audit and challenge AI output adversarially rather than produce first drafts, and reports professional indemnity insurers hardening their AI-governance requirements before renewal.¹⁰ That last point is one practitioner's observation, not survey data. But the direction fits: a supervision policy that exists as configuration is something an insurer can inspect before setting a premium.

None of the paper's prescriptions should feel foreign to lawyers, because courts arrived at most of them long ago. An appellate bench is a multi-agent system the law already trusts: judges form views independently before they confer, dissents are published because disagreement is information, and no rule forces unanimity. A real two-to-one split tells you something a manufactured three-nil cannot. The split-vote finding is the same design instinct with a measurement attached.

The analogy breaks in one place. A bench's unanimity means something because the judges are different people, with different training, different careers, different priors. L-MAD's agents shared one base model, so their unanimity can be a single blind spot speaking in three voices. A panel of agents drawn from one model is not a bench. It is one judge, photocopied three times slightly differently.

That is what the 17-to-24% floor is. Those cases did not fail because the agents reasoned badly. They failed because every copy of the judge was missing the same knowledge, and no amount of conferring supplies what nobody in the room knows.

There is an obvious engineering rejoinder: seat different models on the panel. Whether mixing model families breaks the shared blind spot is untested, and it may well help with shared reasoning failures. What it cannot do is supply knowledge that no base model contains — the firm's matter history, its precedents, its playbooks, the way it papers a deal. The knowledge layer binds either way.

The conclusion is an unfashionable one. Model choice matters at the threshold, which was the first setting; past the threshold, the binding constraint on an AI-native law firm is the knowledge layer its agents can reach. Whatever else these firms compete on, they will compete on how well their knowledge is organised — which was always the binding constraint in law firms. What has changed is that it is now a measurable property of a system.

Five settings, then: the floor, the budget, the protocol, the escalation rule, the knowledge boundary. Checking the market against that list shows two things. Fragments of the discipline are appearing everywhere, and no commercial operator shows you the whole configuration.

Lavern, the open-source 67-agent experiment that preceded Brahe, is the most transparent architecture in the category, because you can read all of it. It gates each agent's output through an evaluator running on a different model. Its debate board records every resolution with evidence weights, confidence scores and a human-escalation recommendation. Its verification pipeline fails closed, so an API error counts as a rejection. And it grounds citations mechanically, string-matching each quote against the source document rather than asking a model to judge. Innanen's own write-up is candid about the limits: making agents truly listen to each other, he writes, is unsolved.¹¹ One observation of my own, not the paper's: Lavern can run locally on small models, and small models are exactly the regime where the paper found debate most dangerous.

Norm lists “supervisory agents” in the use of funds for its Series C, which is supervision sold as a product line.¹² Harvey published the architecture of its internal agent platform, Spectre, in April 2026: durable runs, sandboxed execution, explicit permissions.¹³ Its head of applied research closed an April 2026 research thread on legal agents with a line borrowed from OpenAI's engineers.¹⁴ Garfield runs the same shape with its escalation points fixed from outside — the human sits at the advocacy and the accountability because the regulator put the human there.⁶

The paper's ceiling effect suggests frontier models may gain least from debate, and commercial products run frontier models, so a firm could reasonably decide to run no debates at all. Fine. Three of the five settings go with the panel; two do not. Every agent system, panel or pipeline, still has an escalation path and a knowledge boundary. And a firm that drops the panel has also dropped the split vote, the one free uncertainty signal the evidence found, which leaves it owing a different answer to the same question: what tells you when the machine is unsure?

Now try to find a commercial firm or vendor that publishes its full configuration. I could not. Lavern publishes everything, but Lavern is an open-source experiment with nothing to sell. Garfield's escalation points are public because a regulator set them, not because the firm chose to disclose.

Which yields five questions any client, panel reviewer or insurer can put to a firm that runs agents.

The answers matter less than whether a firm can produce them at all. A firm that cannot say what ends its debates has not chosen a stop rule; it has inherited one from whoever wrote the loop.

Two cautions. Research on multi-agent systems reports that identifying which step caused a failure succeeds only 14 to 53% of the time,⁹ so append-only audit trails matter more than dashboards. And adoption is running well ahead of measured benefit: corporate legal AI adoption doubled from 23% to 52% in a year, while only 7% of legal departments report any reduction in total matter cost.¹⁵

Architecture criticism is cheap. So I built the smallest version I could: a three-agent debate panel for construction law documents. This is a demo — synthetic documents only, not legal advice, and every output is a starting point for human analysis. But every setting on the list is visible in it, and the important ones are enforced in code rather than requested in prompts.

One setting exists only to fail on demand. Because the paper shows what forced consensus costs, the system keeps a forced-consensus mode as a labelled anti-pattern: feed it the same split matter and it manufactures unanimity, on the record, so the destroyed signal is visible rather than theoretical.

The three debaters share one base model, so this panel is the photocopier from the section above. That is exactly why the trust here sits with the grounding gate, the declared knowledge gaps and the escalation rule rather than with unanimity itself. A unanimous verdict in this system means every finding survived a mechanical citation check against the vault. In a demo, “auto-confirmed” means confirmed for a human to read.

On 26 July 2026 I ran two demonstration matters end to end. A payment claim with a fatal endorsement defect came back unanimous on all three issues and was auto-confirmed, every finding pinned to a verbatim statutory quote. An extension-of-time claim with an arguable notice trigger went the other way: the panel was unanimous that the statutory route for the delay costs was open, held the time bar unresolvable on the supplied documents, and the system did what the paper says it should. It stopped, and handed the file to a lawyer with the two rival readings of the notice clock mapped.

None of this needed a research lab: three personas, one paper taken literally, and a few hundred lines of guard code, built in the evenings. A billion-dollar platform and the firms built on it could publish their settings tomorrow.

The most sophisticated thing the panel does is hand the file back. The machines argue, the code counts, and a human signs.

The AI-native firms are probably right about the shape: a lawyer above a panel, rules in between. The paper supplied the first legal-task evidence on what those rules have to do. Between those two claims sits everything that matters — supervision that is real only if it is built, and trustworthy only if it can be inspected.

When a firm becomes software, its professional obligations do not dissolve. They compile. Which is, on balance, good news: a policy compiled into configuration can be read by a client, priced by an insurer and audited by a regulator in a way that a policy carried in a firm's culture never could. The first AI-native firm to publish its settings will make every other firm explain why it hasn't.