CHAPTERS
Legora in brief: collaborative AI workspace for lawyers
Jakob Emmerling introduces Legora’s mission: an AI-native workspace that helps lawyers complete end-to-end legal tasks collaboratively. He sets credibility context with company scale and adoption, then pivots to the core topic—what legal agents can learn from coding agents.
The inflection point: realizing legal agents were lagging coding agents
Jakob describes a key realization from about six months prior: progress in coding assistants (autocomplete → chat → agents) was outpacing other vertical agents. This prompts an investigation into why coding agents are so effective and how to replicate that acceleration in legal workflows.
Parallels between coding and legal work (and knowledge work broadly)
He outlines why legal work resembles software engineering: both are document- and precedent-driven, rely on strict conventions, and require strong review processes. These similarities suggest agent patterns from coding can carry over to legal tasks.
A three-bucket framework: reuse, translate, and invent
Jakob introduces a practical framework for vertical agents: directly reuse proven agent UX/patterns, translate coding patterns into domain equivalents, and invent domain-specific capabilities where no direct analogy exists. This structure guides the rest of the talk’s examples.
Reuse pattern #1: planning mode for long-running tasks
He explains how planning upfront improves outcomes with coding agents and maps cleanly to legal work. Lawyers can co-create a detailed plan with the agent, resolve assumptions early, and then let execution follow the agreed structure.
Reuse pattern #2: approval gates for dangerous tool actions
Jakob highlights tool-call approvals as a universal safety/UX pattern. Just as coding agents should not run unsafe shell commands without consent, legal agents should not perform destructive document actions without explicit user approval.
Translate challenge: why DOCX editing is hard and how Legora first approached it
The talk shifts to document editing, emphasizing that lawyers live in Microsoft Word and that DOCX is structurally complex (zipped XML with metadata). Jakob describes Legora’s earlier multi-model handoff architecture and the coordination problems it created.
Coding-agent pattern applied: the read–edit–verify loop with a DOCX intermediate representation
Comparing with coding agents, Jakob notes most converge on iterative read/edit/verify cycles with simple editing tools and validation steps. Legora translates this into legal by converting DOCX into a flat intermediate representation and enabling an iterative tool loop that the model can self-check.
POC breakthrough: exhaustive paragraph-by-paragraph translation (even on a small model)
A proof-of-concept test—translating a 10-page document paragraph-by-paragraph—demonstrates the new harness’s robustness. The agent catches missed paragraphs via re-reading, completes the task end-to-end, and does so even using a smaller model, validating the approach.
Translate pattern: “linting” for legal documents as a feedback loop
Jakob introduces the idea of static and semi-static verification for legal documents, analogous to ESLint/typecheckers for code. Automated checks can detect broken internal references or structural inconsistencies after edits, and can be extended with LLM-based checks for less mechanical issues.
Inventing domain-specific capability: due diligence over thousands of documents
He explains why some requirements are unique to legal practice, focusing on due diligence—reviewing massive contract sets during transactions. Solving this requires specialized workflows beyond generic agent patterns, including scalable structured extraction and review surfaces.
Tableau Review: structured extraction and collaborative verification surface
Legora’s Tableau Review provides a grid interface where rows are documents and columns are extracted fields, enabling filtering, triage, and targeted follow-up. The agent can use the same feature as a human to create structured views across large corpora and surface key issues.
Live demo #1: planning and executing edits across employment agreements and policies
Jakob demonstrates the agent creating a plan to add an extra Christmas vacation week, identifying impacted agreements and the HR policy. The agent then executes via the read–edit–verify loop, produces redlines with correct formatting, and even drafts an employee announcement memo.
Live demo #2: mass document review with Tableau Review (categories, parties, red flags, sorting)
He demonstrates a due-diligence-style task over ~100 documents: creating a tabular review, extracting document categories, parties, and red flags, and enabling citation-based verification. The agent then uses the results to organize documents (e.g., moving employment agreements into a folder).
Why coding is ahead—and the closing takeaway: keep stealing from coding agents
Jakob closes by reflecting on why coding leads in AI adoption (engineers’ tool openness, compounding benefits of solving coding). The practical conclusion is a repeatable strategy for any vertical: monitor coding agents, reuse what fits, translate patterns, and invent the domain-specific remainder.
