Skip to content
ClaudeClaude

What legal agents inherit from coding agents: Lessons from Legora

Three patterns shaped Legora's legal-AI agent: what they could reuse from coding agents, what they had to translate, and what they had to invent. Hear Staff Software Engineer Jakob Emmerling share how Legora rebuilt document editing, linting, and bulk review around coding-agent principles and see the agent live.

May 21, 202628mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

How Legora borrows coding-agent patterns to build legal agents

  1. Legora found legal agents lagged behind coding agents and analyzed why coding agent UX and tool harnesses work so well.
  2. They categorize learnings into three buckets: reuse one-to-one (planning, approvals), translate patterns (document editing loops, linting), and invent domain-specific capabilities (citations, due diligence workflows).
  3. A key breakthrough was replacing a brittle multi-model handoff system for DOCX editing with a coding-agent-like read–edit–verify loop using a text-based intermediate representation.
  4. Legora applies “linting” concepts to contracts (e.g., validating cross-references) to create a feedback loop similar to type-checking and CI in software.
  5. Live demos show planning-first execution for updating many employment agreements and using a tabular due-diligence review surface to extract, verify, flag red flags, and organize documents at scale.

IDEAS WORTH REMEMBERING

5 ideas

Adopt a reuse–translate–invent roadmap to build vertical agents faster.

Treat coding agents as the leading indicator: copy universal interaction patterns, adapt analogous workflows, then focus engineering effort on the irreducibly domain-specific last mile.

Planning-first interaction is a universal pattern for long-running agent work.

Iterating on a plan up front collects context and locks decisions so execution becomes more deterministic—mirroring how engineers use coding agents for larger tasks and how lawyers prefer structured work breakdowns.

Human approval gates should map to “dangerous actions,” not just model confidence.

Like approving shell commands in coding tools, legal agents should require explicit confirmation before irreversible or high-risk actions (e.g., deleting/moving client documents or bulk changes).

Avoid brittle multi-model handoffs when the agent’s capabilities expand.

Legora’s earlier DOCX approach used multiple specialized LLM calls with differing tools/contexts, creating integration and “missing tool” failures as the system grew; a unified loop reduces these coordination errors.

Make documents editable like code by introducing an agent-friendly representation.

DOCX is noisy (zipped XML), so Legora flattens it into a text-based intermediate representation and provides read/edit tools, enabling a coding-agent-style read–edit–verify loop with better exhaustiveness.

WORDS WORTH SAVING

5 quotes

My mental model is kind of that you wanna have the model almost feel like it's inside a coding agent harness, and it just does a legal task.

Jakob Emmerling

Whenever they ship something new, you steal the thing and benefit from it.

Jakob Emmerling

We see it's quite different. Basically, like, all the coding agents out there work in a way where they, they just read, edit, and verify things in a loop.

Jakob Emmerling

At the end, we open up the document and, uh, everything was, was translated.

Jakob Emmerling

There's basically three buckets how you can, how you can learn from coding agen-agents, as we found.

Jakob Emmerling

Why coding agents advanced faster than other vertical agentsThree-bucket framework: reuse, translate, inventPlanning mode and human-in-the-loop approvalsDOCX editing via intermediate representation and edit loopsReducing handoff brittleness from multi-model orchestrationLegal “linting” for mechanical contract consistency checksDue diligence at scale with tabular extraction, citations, and verification workflow

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.