CHAPTERS
- 0:15 – 1:01
Legora’s mission and the shift in how they build legal agents
Jakob introduces Legora’s collaborative AI workspace for lawyers and explains that a key realization six months prior forced them to rethink their agent architecture. The talk frames legal agents through the lens of what has already worked in coding agents.
- •Legora builds an end-to-end AI workspace for legal tasks with broad adoption
- •A recent internal realization prompted a change in their agent approach
- •Goal of the talk: extract lessons from coding agents for legal work
- 1:01 – 2:01
Why coding agents surged ahead—and why other vertical agents lagged
The talk highlights the rapid evolution of AI for coding from autocomplete to full agents, contrasted with slower progress in domains like legal. Jakob sets up the central question: what makes coding such fertile ground for powerful agents?
- •Coding AI progressed quickly through multiple capability phases
- •Legal and other verticals were notably behind until recently
- •Motivation: identify the ingredients behind coding-agent acceleration
- 2:01 – 3:03
Parallels between coding and legal work: conventions, text artifacts, and review culture
Jakob maps similarities between software engineering and legal practice that make agent patterns transferable. Both domains rely on prior artifacts, strict conventions, and rigorous human review workflows.
- •Both domains are heavily based on reusing prior work/templates
- •Work outputs are largely text-based documents
- •Strong conventions within firms/organizations guide correctness
- •Review culture parallels PR review vs partner sign-off
- 3:03 – 4:33
A 3-bucket framework: reuse, translate, and invent
Legora organizes learnings from coding agents into three categories: things you can reuse directly, patterns you can translate, and domain-specific capabilities you must invent. This becomes the guiding structure for the rest of the talk.
- •Reuse one-to-one: established agent UX and control patterns
- •Translate: adapt coding patterns to legal equivalents
- •Invent: create domain-specific tools like citations and large-doc workflows
- 4:33 – 5:34
Reusable pattern #1: Planning mode as the default for long-running work
Jakob explains why planning-first workflows make agents more effective and controllable for complex tasks. Legora applies the same plan-iterate-execute flow for legal matters to reduce ambiguity and agent decision-making mid-flight.
- •Planning helps gather context and align on assumptions early
- •Reduces agent autonomy at execution time by front-loading decisions
- •Directly mirrors how users collaborate with coding agents on big tasks
- •Legal users can iterate on plans before execution
- 5:34 – 6:37
Reusable pattern #2: Human-in-the-loop approvals for risky actions
Tool-call approvals in coding agents (e.g., shell commands) translate cleanly to legal workflows where destructive or sensitive actions must be gated. Legora adopts this UX to avoid re-learning safety interaction design from scratch.
- •Certain actions require explicit user approval for safety
- •Legal analog: prevent accidental deletion/modification of client documents
- •Reuse of established coding-agent UX avoids costly iteration cycles
- 6:37 – 7:38
Translating coding patterns to Word: why DOCX editing is uniquely hard
Jakob introduces the challenge: lawyers live in Microsoft Word, but DOCX isn’t plain text—it’s a zip of XML with complex structure. This makes naive LLM editing approaches brittle and difficult to execute exhaustively.
- •Legal workflows depend heavily on drafting/redlining in Word
- •DOCX complexity (XML/metadata) complicates edits vs Markdown/text
- •Agent must handle formatting, structure, and long documents reliably
- 7:38 – 9:09
Legora’s earlier DOCX solution: handoffs, marker-based edits, and scaling pains
The initial approach split the task across multiple models: one reasoned about what to edit, then others generated full edits per location. While it improved exhaustiveness, it created fragmentation, inconsistent context, and tool mismatch issues as the system grew.
- •Top-level agent delegated to a reasoning model to identify edit points
- •Separate calls generated detailed edits (style-aware) per marker
- •Reduced token use and laziness, but introduced handoff complexity
- •Tooling/context mismatches emerged as the agent gained capabilities
- 9:09 – 10:10
The coding-agent edit loop: read → edit → verify, and why it works better
Jakob contrasts their earlier system with the common coding-agent architecture: iterative reading, patching, and verification (often with tests/linters). Seeing coding agents make many “surgical edits” reliably inspired Legora to adopt the same loop for legal docs.
- •Coding agents converge on iterative loops rather than multi-model handoffs
- •Line/patch editing plus verification yields reliable multi-location changes
- •Experiments on large documents showed better exhaustiveness behavior
- •Motivated redesign of legal document editing to match the loop
- 10:10 – 11:12
New DOCX approach: intermediate representation + editing tools + verification
Legora rebuilt Word editing to feel like a coding harness: convert DOCX into a flat text-like intermediate representation, then apply editing tools in an iterative loop. This allows the model to re-read its own changes and correct omissions over time.
- •DOCX is transformed into a flat intermediate representation for interaction
- •Agent uses read/edit tools repeatedly and checks its work
- •Loop enables self-correction and reduces missed edits in long documents
- •Tool design intentionally mirrors coding-agent ergonomics
- 11:12 – 13:14
POC breakthrough: paragraph-by-paragraph translation succeeds—even on a small model
A quick proof of concept tested exhaustive translation of a long document—previously a failure case—and it completed successfully by looping, re-reading, and fixing missed paragraphs. The surprising success on a weaker model (Haiku) validated the harness-first hypothesis.
- •Test case: translate a 10-page document paragraph-by-paragraph
- •Agent looped, revisited missed sections, and completed exhaustively
- •Worked even using Haiku, implying harness/tooling mattered greatly
- •Reinforced idea: “make the model feel like it’s in a coding agent harness”
- 13:14 – 14:45
Legal linting: ESLint-like feedback loops for documents and cross-references
Jakob proposes ‘linting’ legal documents to catch mechanical issues—especially when edits break references or structure. Adding deterministic (and optionally LLM-assisted) checks gives agents the same kind of corrective signal that linters and type checkers provide in software.
- •Static checks help agents avoid mechanical/document-structure mistakes
- •Example: detect broken cross-references after edits
- •Feedback loop prompts targeted fixes similar to code lint/test cycles
- •Can be extended with LLM-based lint rules beyond purely mechanical checks
- 14:45 – 17:16
What must be invented: due diligence over thousands of documents (Tableau Review)
Some legal workflows have no direct coding equivalent—like due diligence across massive contract sets—so Legora built domain-specific surfaces. Tableau Review enables structured extraction into a grid, filtering by red flags/parties, and scalable review workflows.
- •Due diligence requires reviewing thousands of contracts/documents
- •Tableau Review: grid interface where rows are documents and columns are extracted fields
- •Enables filtering, triage, and targeted deep dives on risk areas
- •Represents the domain-specific “last mile” of an effective legal agent
- 17:16 – 21:50
Live demo #1: planning and executing edits across employment agreements and policies
Jakob demonstrates Legora’s agent planning a change (extra Christmas vacation) by discovering relevant agreements and policies, then executing edits using the iterative edit loop. Users can review redlines and approve changes before writing back.
- •Agent first explores project files and produces a step-by-step plan
- •Execution applies the read/edit/verify loop across multiple documents
- •Edits are staged for review with redlines and formatting preserved
- •Agent also proposes ancillary outputs (e.g., announcement memo)
- 21:50 – 27:25
Live demo #2: mass review with Tableau Review—citations, verification, and filing
The second demo shows the agent creating a Tableau Review for ~100 documents, extracting categories/parties/red flags, and enabling interactive verification with highlights/citations. The agent then performs an organizational action: moving all employment agreements into a folder.
- •Agent initiates a structured review and creates a tabular extraction grid
- •Users can inspect reasoning, click citations, and see source highlights
- •Verification status supports collaboration (associate review, senior oversight)
- •Agent completes workflow by classifying and filing documents into folders
- 27:25 – 28:42
Closing: coding is ahead—so keep stealing the patterns
Jakob closes by reflecting on why coding leads AI adoption (tool openness, leverage), but argues the reason doesn’t matter operationally. The practical framework is to continuously reuse, translate, and invent as coding agents ship new capabilities.
- •Possible reasons coding leads: willingness to adopt tools and high leverage
- •Vertical builders can ignore ‘why’ and focus on harvesting proven patterns
- •Framework recap: reuse what fits, translate what’s analogous, invent the domain-specific remainder
- •Ongoing strategy: track coding-agent releases and adapt their innovations
