I've been quiet as I've been on holiday. ☀️🍹 As I sat on a lounger, I found myself reading the source code of MikeOSS. 🤓

If you've missed it, MikeOSS is ex-Latham lawyer Will Chen's take on an open-source competitor to Harvey and Legora – the two heavyweights of legal AI tooling.

Chen 'vibe-coded' MikeOSS, apparently creating the first version in two weeks. Looking at the source code, it does reveal some of the challenges of this approach, not least with security. Nevertheless it's an impressive endeavour, and the benefit of being open-source is that the community can help improve it over time.

What strikes me most with these state of the art of these tools is how similar many of them are:

• Chatbot assistant for legal research ✅

• Left-hand 'drawer' for session management (chats, tasks, topics, workflows, agents) ✅

• Large scale document review with tabular reporting ✅

• Document redlining and Microsoft Word integration ✅

• Connectors to data sources - web, internal documents, legal databases etc. ✅

This is the case, not only for Legal industry specific platforms, but also with the Claude for Legal plugin, which makes the Anthropic's general purpose desktop AI tool Claude suitable for legal workloads. (Helen Fan's take on LinkedIn)

Of course, these software packages are at varying states of maturity. There's a lot to consider in whether HarGora or Claude for Legal is superior for any particular legal team or task. Perhaps UK home-grown Wordsmith ($70m Series B round announced this month) also deserves a shout-out... their demo at CraftyFest the week-before-last was impressive. I'm at LegalTechTalk at the minute and I'm enjoying exploring the best the market has to offer.

AI = the new IDE for lawyers?

Given their high-level similarity, perhaps we can say that there is an emerging class of AI software for lawyers?

It has bugged me that most lawyers do not routinely use domain-specific software... our own 'IDEs' if you will. Instead we use general-purpose office tools like Word and PowerPoint. These tools tend to have decades of crusty code - which makes them brittle against innovation. They are the default through adoption (the network effect) rather than brilliance.

I've been chatting with the folk from Vesence. We observed that lawyers use these office programs - especially Microsoft Word - to the limit. They tell me they've had to go really deep on the MS Office file formats to teach their model how lawyers work.

E.g. the difference between a footnote, an endnote, and (no-doubt) some weird law-firm-specific macro that formats them for us.

Many other comparable knowledge-workers use specialist software packages: developers, architects, data analysts, doctors... Perhaps AI is finally giving lawyers our own toolset. It does mean we need to get good at using it. Not just as a basic chat-bot, but in-depth, detailed working knowledge of the LLM-first system.

As I dig more into Claude I can see how much functionality there is waiting to be explored. I can also see how the providers are developing their tools slowly away from AI chat interfaces and towards being platforms, especially for workflows and agents. Perhaps this won't be Legal-specific after all. But it will be different from the word-processor era.

My own agent

It would have been ideal if my agent paralegal — newly christened Reginald — had been autonomous on holiday, leaving me to oversee from the beach. Instead, Reginald still needs a fair bit of manual operation, and I confess I rather neglected it in favour of the good weather.

I am now working on reducing its reliance on me. Initially through Claude's scheduled tasks functionality (well worth a look if you are playing along), and ultimately via multiple, orchestrated, agents. I am testing CrewAI for this but suggestions welcome. Reginald will then run without me needing to open my laptop. Blog post coming soon.

Until now, I've still not needed to write any actual software code for Reginald. One good friend did point out that prompting is essentially programming: having the AI create and execute small-scale, disposable software programs. His point: doing so with natural language is not so dissimilar to doing so with code (if you want to geek out on this, this article from Towards Data Science is helpful).

Perhaps then, we're all programmers now. But when /vacation-mode is on we should probably stick to fiction. 📚

Around the web

• Talking of which, in fiction, AI over-explains themes and favours "tidy, single-track plots". Each model also has its own tells: "for example, Claude produces notably flat event escalation, GPT over-indexes on dream sequences, and Gemini defaults to external character description". When it comes to legal drafting, will we be able to identify the model behind the mark-up? Just as we used to identify the law firm behind the precedent?

• Chance might be a fine thing; perhaps we won't have the time. Lawyer Sebastian Foerste on the Jevons paradox:

"Make a resource cheaper and the world consumes far more of it. [...] The photocopier multiplied discovery, and email did the same to correspondence. [...] AI now makes legal output close to free. Follow the pattern to its end: the amount of law in the world is about to explode. Contracts that nobody ever reviewed will get reviewed. A claim worth 80 euros becomes worth enforcing..."

• Talking of a step-change, I didn't get to test Reginald on Anthropic's latest model, Fable 5 (System Card), before that model was pulled citing national security concerns. The model is reported to be a real improvement for long-running agentic processes. Current models can struggle after a while, especially with large or complex 'context' (input / base materials); hopefully by the time Reginald's work grows to a size that thwarts current models, Fable 5 will be back.

Useful links

• Want a shortcut to your own agent or workflow? Download a community-developed skill from LawVe.ai. For example: Need to review outside counsel bills? There's a skill for that. Great for use with care. Read the skill's source-code first, it's expressed in natural language after all.

• Want to skip straight to building an app? I liked the look of Fliplet at LegalTechTalk, a no-code software / app building platform which has gone long on AI smarts. I hope to have a play with it over the next few weeks, as well as others in this class of AI-native 'app builders', like Loveable, and Base 44. If you want to use AI to create apps quickly for your own use, I think they'll be well worth exploring.