Most of the legal tech covered by The Legal Wire fits a recognisable shape: a Series A, a launch date, a careful answer to questions about defensibility and procurement. Lavern, the project Antti Innanen released as open source on the 20th of May, is not that. It is a multi-agent legal system, 155,000 lines of code, sixty-seven agent prompts, nine workflows and an orchestration engine, given away under Apache 2.0 with an invitation to do almost anything with it. Look inside, take what you like, build a company with it (or don’t), and if you never mention him, that is, he says, more or less the point.
Antti is a tech lawyer by training. He has worked in larger firms and smaller ones, co-founded one of his own, and spent recent years pulled into legal design and then into AI, trying to bridge two worlds that do not naturally talk. Legal design is about making law accessible, well-shaped, sometimes even fun. AI in law has mostly been about doing the same old work faster and at greater volume, which is useful and, in his slightly dry framing, not necessarily interesting. Lavern is what happens when someone interested in the first set of questions sits down to build with the tools of the second.
Lavern was never planned. He started with about ten agents, kept adding, and eventually realised he had built something larger than himself. There were conversations about selling the IP, about acqui-hires, about VC money; but none of them felt right, and so the project went open instead.
He is candid that this is not pure altruism. He wants people to play with it, break it, find the bits worth lifting out and shipping as something of their own. He would rather be the person whose repo a hundred other projects forked than the person who held the IP and never quite shipped.
What is in the box
Lavern’s central conceit is that it resembles a small law firm whose staff are AI agents. They have names, faces, specialisms, personalities and skill ratings, the kind of bench you might recognise from a sports video game. You can use the pre-loaded team, swap people in and out, build your own agents, or clone yourself as one, which is the sort of feature you assume is a joke until you read the documentation.
Underneath the cast of characters, the architecture is more serious than the framing suggests. An intake mode gathers context through an interview agent. Specialists discuss the matter at a debate board, where findings cite evidence and challenges cite counter-evidence. An orchestrator weighs both and writes the document. A ten-pass verification loop then checks references, abbreviations, numbering and structure, sending work back when it does not pass. The system is built around loops rather than linear pipelines, which is how legal work actually behaves: a draft is rarely a draft so much as one orbit through a circuit of checks.
A second mode, Clawern, runs autonomously on a local machine, watching a folder and checking documents on a thirty-minute heartbeat. Local models are not yet as capable as frontier ones, perhaps two years behind, but they are always on, cost nothing per query, and your documents stay on your computer rather than being shipped to a cloud in California. Antti thinks this is the more interesting horizon: the local model handles the repetitive, low-stakes checking work, and the system only escalates to a frontier model when the question actually deserves it. For a profession that spends a great deal of time checking whether something has changed, that’s a pretty useful architecture to explore.
Sixty-seven agents and the problem with crowded rooms
The sixty-seven-agent count gets attention, which Antti finds slightly funny. He did not arrive at the number through any principled architectural decision. He started with ten, thought an environmental agent would be fun, then a risk-pricing one, then another, and simply kept going. “It became like an agentic HR department,” he says. About ten to twenty agents do most of the real work; the rest are options on the bench.
More importantly, he is open that more agents is not better. Past around ten, debates degrade in a way anyone who has sat in a real meeting will recognise: someone dominates, others go quiet, agents end up siloed in their own reasoning, or swing wildly to the opposite position when challenged, not because the challenge was strong but because it was new.
The sweet spot for Antti is “when [he’s] doing eight agents, max”. The system ships with workflow templates of different sizes for exactly this reason. Getting AI agents to genuinely listen to one another, he says plainly, remains an unsolved problem.
TLW: You’ve been unusually candid that getting agents to listen to each other is unsolved. When two well-matched agents reach incompatible conclusions and refuse to budge, what does the orchestrator actually do in that moment and what does that teach you about how multi-agent systems should be designed?
Antti: “Agents don’t always agree. And the orchestrator does not force consensus. It records the disagreement.
Each agent’s finding stays on the debate board with its citation. Then there is a synthesis step, where the AI either reconciles the two views with the reasoning intact, or surfaces the split in the deliverable for a human to settle at the next gate.
What I learned building this is that the wrong question is “how do we get the agents to agree.” Agreement at any cost produces the worst kind of multi-agent failure.
But functioning agentic debate not easy to design. I have tried to handle it with the debate board, orchestrator and then escalation to a human.”
What is genuinely legal about it
On the question of whether Lavern is really a legal tool or a general-purpose multi-agent framework with legal training data laid on top, Antti explains that nothing in the architecture prevents the system from working in other domains. In fact, people have already mentioned strategy versions of it.
That said, what makes Lavern legal is essentially two things: the training materials and prompts inside the agents, and the context-gathering and checking loops on either end.
This is where his legal-design instincts surface. Legal work, he says, is a context game. Most of the value of a talented lawyer sits in extracting good context from a client who does not know exactly what they want, cannot quite name it, or is using the wrong vocabulary entirely. Lavern’s intake agents are built to do that systematically, and the ten-pass loop at the other end keeps legal-specific things, references, abbreviations, numbering, from quietly going wrong. A single pass produces a draft; a loop produces something you could send.
There is a ‘less obvious’ point underneath this, and it is the one that distinguishes him from most of the legal AI conversation. He does not want Lavern to be too legal. The promise of an agentic system, he argues, is precisely that it lets you bring perspectives other than the purely legal one into the room: a plain-language agent, a designer agent, a client-perspective agent. “You can look at a legal problem from many different angles, and usually lawyers look at it from a legal point of view,” he says.
TLW: You’ve described Lavern’s edge as letting non-legal perspectives such as design, plain language, the client’s point of view sit inside an agentic system. In a concrete piece of work, what would a reader actually see that they wouldn’t get from some of the more well known legal tech tools, and what is the legal industry losing by not building this way?
Antti: “Click the plain-language agent and it will flag a spot where a non-lawyer would stumble. It is almost like the lawyer’s draft is still in there, but it has been challenged by people who are not lawyers.
Think about the legal tech tools on the market right now. How many are building toward making law more accessible, more understandable, or even fun? Most are focused on efficiency and completeness.
But making something understandable and useful takes real skill. It requires deep understanding. And it could be a genuine moat for legal AI.”
Antti is, on balance, more generous about the Hargoras than the framing of his project might suggest. He thinks the products are good, that the value sits largely in the user experience and workflow design, and that the companies have access to some of the best legal minds in the world. He doubts they are demonstrably better at producing legal answers than the foundation models underneath them, but is careful to say that could change as they listen to their customers.
What he doesn’t resonate with is the worldview the category projects. “I don’t like the cigars and the marble floors and wooden panels,” he says, with the casual exasperation of someone who has watched too many polished commercials. The complaint is not really about the marketing. It is that the imagery imports an old and slightly tired idea of what law should look like into a moment that could be used to make law look like something else. “I’m not critical of the products,” he says. “I’m critical of what kind legal tools we are building.”
The unusual practice of admitting it doesn’t fully work
There is a section on the Lavern website titled, plainly, “What We’re Still Learning.” In it, Antti writes that the agent debate at the heart of his system is genuinely difficult to get right, that more agents do not always produce better results, that the autonomous Clawern mode is “the biggest promise and the scariest thing,” and that even the law-firm analogy he has built the whole project around may be a constraint he eventually has to discard. It is the sort of section most legal tech founders would have a marketing team write out of the site.
Antti has done the opposite.
He is consistent about this in person, too. “AI is not that good at listening,” he said on the call, almost cheerfully, as if it were obvious. He calls the agentic-debate problem the hardest thing about the system, says local-model orchestration is the second hardest, says the tool probably might not replace any of the existing ones, and may not even work properly yet, and that none of this seems to bother him very much.
TLW: On the Lavern site you write openly that agent debate is unsolved, that the law-firm analogy may eventually need to be dropped, and that autonomy is the part that scares you most. Most founders would never write that page. Why have you chosen to be that candid, and is there a cost to it that you have already noticed, commercially or otherwise?
Antti: “A lot of legal tech discussion is just what problems are solved. Maybe a new feature pops up, some tools are winning the evals. Less is talked about what we haven’t yet solved or where we screwed up.
Open source is fun because it kind of gives you a concrete thing you can use as a base for this discussion. Suddenly the questions are not so theoretical anymore. You have to decide what the agent actually does!
I am candid because I have to be. The code is out there, for everyone to look at. Better to start with the truth and let people decide from there.”
TLW: You’ve written that the law-firm metaphor has been generative. It produced the soul, the debate board, the partner consultations, the retainer model, but that you may eventually let it go and find out what the system wants to become when freed from inherited shapes. What does that freed version look like, even in early outline, and what is currently holding the analogy in place?
Antti: “The law firm analogy is a fun one, but it is still skeuomorphic. It imitates the old world.
I have been accused of building a “vegan burger” that looks like meat.
And that is kind of true. But people may need something familiar first to wrap their heads around the concept. The next iterations might be crazier.
In the future I think we will see something similar to what happened with chess engines. We do not understand the moves until thirty moves later. I do not think agents are going to be sending NDAs to each other or doing contracts the human way.
It will probably be like an alien language to us, and we will have to rely on AI to explain it back to us. Exciting and scary at the same time.”
The invitation
If there is a single thing Antti wants from this article, it is small and unusually concrete. He would like someone to click the GitHub link. Maybe, he allows, it will be the first GitHub link they have ever clicked. Then they open the repo, point Claude or ChatGPT at it, and ask the AI to explain what is going on inside an ambitious legal tech project. Harvey, Legora and the other big ones are, he says, black boxes. Lavern is the opposite of that on purpose, sitting there ready to be poked at, criticised, forked, remixed, or ignored.
TLW: If a reader of this piece closes it, opens the repo for the first time, and points their AI tool at it the way you suggest, what would you most want them to come away with? A specific module to lift, a way of thinking about agents, or something else entirely? And what would indicate to you, a year from now, that the open-source bet had worked?
Antti: “I hope this gets people thinking differently about what legal AI could be, and what kind of worldview it could project.
Legal AI does not have to be a chatbot. It does not have to deliver legalese. It does not have to look at law from a purely legal angle. And non-technical people can contribute too. We are not bound to the old ways of building.
I am already very happy with the open-source decision. I see people discussing the tool, building on top of it, and getting inspired. That is what I wanted from it.”
That is an unusual ambition for a launch piece, and the reason Lavern is hard to file neatly. It is not really a product story, and not really an open-source manifesto either, though it has elements of one. It is closer to the work of a curious builder who happens to be a lawyer, has spent years thinking about how to make law more human, and has put a working, flawed, deliberately legible piece of agentic legal infrastructure into the public domain to see what people will do with it.
So, if you have ever idly thought about what an agent for legal aid, or plain language, or design review might look like, the repo is right there, the licence is permissive, and the building costs, as Antti points out, are approaching zero.
