Building Legal Agents When You Can't See the Data - Jacob Lauritzen, CTO of Legora

What it claims

The page is show notes for a Latent Space episode (about 66 minutes) with Jacob Lauritzen, CTO of Legora. The transcript is not included. The notes add several specifics beyond the summary:

• Lauritzen estimates that current models already do about 60% of legal work well, and that the rest needs a smarter model.
• Legora has 100 legal engineers and a separate evals team.
• Three engineers and three lawyers build the simulated test environments. One of them is a 10,000-file data room.
• Paraphrased traces are used to narrow agent failures down to 10 or 20 failure types.
• For drafting, Lauritzen finds Fable and Opus too chatty and prefers Astra's concise writing.
• He argues that drafting a long contract is a search problem.
• He advises against default embeddings.
• Legora resisted adding a model picker for years before adding one.

How it works

The notes name the topics but give no mechanism. The chapter list shows what the episode covers:

• Search and retrieval, including late interaction (54:05).
• Post-training open-weight models for legal research (53:05).
• Model arbitrage with open-weight models (30:04).
• MapReduce versus one large context window for a million-token contract (58:19).
• A citation system aimed at hallucinations (57:03).
• Whether lawyers actually check citations (57:41).
• How agents create lists and how a list can start another agent (39:11).

Tabular Review lets a lawyer review 10,000 contracts in one table. The episode covers what the first 10 rows show (35:51). The intake bot runs in Slack (37:53).

Evidence

Citation accuracy went from 70% to 99% after one model release, with no change on Legora's side. The notes do not name the model, the release or the eval setup.

Limits

These are episode notes only. Every figure is a claim from the interview, and none comes with a method. The notes do not give:

• the reason Legora does not use ElevenLabs,
• the embeddings that Lauritzen recommends instead of default ones,
• the answer on MapReduce versus a long context,
• what the late-interaction setup looks like.

A reader who wants those has to listen to the episode at the timestamps above. Lauritzen's engineering headcount also changed: an earlier slide argued that 20 engineers would be enough forever, and Legora now has 100.