CHAPTERS
Legora’s mission: collaborative AI workspace for end-to-end legal tasks
Jakob Emmerling introduces Legora and frames the talk around building agents that can perform real legal work inside a lawyer’s workflow. He sets context on Legora’s scale and why agent design choices matter in production legal environments.
Why coding agents raced ahead—and what that means for other verticals
The talk uses the familiar progression from autocomplete → chat → agents → background agents to highlight how quickly coding has advanced. Jakob explains that legal (and many other knowledge-work verticals) were noticeably behind, prompting investigation into why coding is special.
Parallels between coding and legal work: text, conventions, and review culture
Jakob outlines why legal work is structurally similar to software engineering. Both domains revolve around text artifacts, reuse of prior work, organizational conventions, and rigorous review/approval processes.
A 3-bucket framework: reuse, translate, and invent
Legora organizes learnings from coding agents into three categories. Some mechanisms can be reused directly, others require translation of patterns, and some must be invented due to domain-specific requirements like citations and large document sets.
Reuse example: planning mode as the default collaboration UX
Jakob describes how structured planning upfront improves long-running agent tasks in coding and maps cleanly to legal workflows. Lawyers can iterate on assumptions and decisions first, then let the agent execute reliably.
Reuse example: approvals for risky tool calls and actions
He explains the importance of keeping humans in the loop for dangerous operations. The same safety/UX pattern used in coding (approve shell commands) applies to legal (prevent destructive actions on client documents).
Translate example (part 1): why DOCX editing is hard and the initial multi-model approach
Jakob dives into document editing—core to lawyers who live in Microsoft Word. He explains why DOCX is complex and how Legora initially used a multi-step, multi-model pipeline to achieve exhaustive edits, but encountered scaling and coordination problems.
Translate example (part 2): adopting the coding-agent read/edit/verify loop for DOCX via an intermediate representation
Legora shifts to the coding-agent paradigm: iterative reading, editing, and verifying in a loop. They flatten DOCX into a text-like intermediate representation and build edit tools over it, enabling the same robust iterative behavior seen in codebases.
POC breakthrough: exhaustive translation succeeded even on a smaller model
A quick, informal proof-of-concept test validated the new approach. Translating a 10-page document paragraph-by-paragraph worked end-to-end, with the agent re-reading and correcting missed sections—demonstrating the power of harness/tool design.
Translate example: “linting” for legal documents to create feedback loops
Jakob introduces the idea of ESLint-like checks for contracts and legal drafts. Static and semi-static validators can catch broken references and other mechanical issues, giving agents a structured feedback loop similar to compilers and linters in software.
Invent example: due diligence at scale with Tableau Review (mass document analysis)
Some needs are uniquely legal, like reviewing thousands of contracts during M&A due diligence. Legora equips its agent with a domain-native tool (Tableau Review) to extract structured data across large document sets and help lawyers triage and investigate.
Live demo (part 1): planning and executing edits across employment agreements
Jakob demonstrates the agent planning a change—adding an extra week of Christmas vacation—then executing across multiple documents. The system copies files to a review space, performs iterative edits, and produces redlines for human review.
Live demo (part 2): due diligence workflow—categorization, red flags, citations, and foldering
A second demo shows the agent creating a Tableau Review over ~100 documents, extracting categories/parties/red flags, and enabling verification with citations and highlights. The agent then performs organizational actions like moving employment agreements into a dedicated folder.
Closing: why coding is ahead and the “steal from coding agents” playbook
Jakob closes by reflecting on why coding leads in adoption—engineer tool appetite and compounding leverage—but emphasizes that the exact cause matters less than the opportunity. The core message: continually reuse, translate, and invent as coding agents ship new patterns.
