Deep research agent decides to use tools, does it over several rounds. Reads, searches again, just as a lawyer would.
Writes up a note, in the exact style I want, with direct links and paragraph citations:
Product comments:
- The output, the research note, should be so close in style in style to what the lawyer wants that they can copy and paste relevant parts, with minimal editing.
- Human verification is still crucial. Direct links to cases, and exact paragraph references, must be provided so that the lawyer can verify the machine's work ASAP.
- Mini-apps like these can also be offered as an API, to plug in to agents which are more general purpose.
Technical observations:
- Early 2025 to early 2026, the overall inference costs of the LLM agent fell by, ballpark, 4x. (Gemini 2.5 Pro -> Gemini 3 Flash)
- Without the big drop in inference costs, this deep research would be too expensive to run as a personal project.
- The LLM agent itself is extremely simple to program. A loop with around 5 tools.
I wrote this app for exactly 1 user - myself. It's never been easier to build custom, personal apps. Python is basically English. AI can write all the boilerplate code for you.
Professional legaltech service providers are still absolutely necessary though. For instance, this homemade deep research app does not yet do sherpardization, i.e. flagging whether cases have been overturned or disapproved of. An LLM can have a good try at automated sherpardization. But it would be expensive for a personal project, and still not as reliable as paying humans to do it assiduously.