Agents need more than a chat - Jacob Lauritzen, CTO Legora

Jacob Lauritzen, CTO of Legora — a legal tech platform serving over 1,000 customers across 50 markets and described as the fastest growing legal tech startup in history — argues that complex, long-running agents fundamentally require richer collaboration interfaces than chat. The talk is structured around a central problem: the familiar failure mode of a 30-minute agentic run that produces a contract, where the user spots an error in Clause 3, attempts a correction, and triggers context compaction ("context rot"), resulting in a revised output that has silently changed far more than requested.

Lauritzen frames the core shift in AI economics: doing work is now extremely cheap, but planning work and reviewing work have become the new bottleneck. This reframes the agent-human collaboration problem — the challenge is no longer just getting a task done but enabling humans to effectively steer and verify agent output across planning, execution, and review stages.

He introduces the "verifier's rule" (attributed to Jason) — a principle stating that if a task is solvable and easy to verify, AI will solve it. He extends this from foundational model training to agent loops: if a task can be made verifiable, an agent can run in a loop and self-correct. Different industries sit at different points on this solvability/verifiability spectrum. In legal specifically: checking contract definitions is easy to verify and easy to solve; drafting a contract is easy to solve but nearly impossible to verify (the true verification only occurs when a judge rules on it in court); and litigation strategy is both hard to solve and impossible to verify objectively.

To move tasks toward the verifiable end of the spectrum, Lauritzen identifies four mechanisms: (1) task reframing — e.g., adding browser access and test-driven development to a coding task converts it into a verifiable loop; (2) proxy verification — comparing new contracts against known-good "golden contracts" as a stand-in for true legal verification; (3) task decomposition — separating easy-to-verify sub-tasks (formatting, definition linting, clause checking) from judgment-heavy sub-tasks (risk profile, negotiation stance) that remain human-owned; and (4) guardrails — constraining the agent's action space (specific files, directories, websites) to increase predictability and trust, analogous to Claude Code's permission tiers.

On the control side, Lauritzen maps agent work as a DAG of sub-tasks and explains how most current interfaces only let humans inject judgment at the root node. He evaluates three approaches: pure planning (aligning upfront, then silence — inefficient and unable to anticipate edge cases discovered mid-run); skills (encoding human judgment directly into workflow nodes, enabling contingency handling and progressive discovery — described as "really, really, really good"); and elicitation (having the agent surface uncertainty to the user mid-run, paired with a decision log that lets the agent unblock itself while preserving human override). He recommends combining skills and elicitation, and specifically cautions that agents should never block on uncertainty — they should make a decision, log it, and continue.

The central thesis culminates in a rejection of chat as the primary collaboration interface for complex agents. Chat is "one-dimensional" and "low bandwidth" — it collapses a tree-structured work process into a linear thread, strips context, and makes reviewing agent decisions nearly impossible at scale. Lauritzen proposes instead that humans and agents collaborate in "high-bandwidth artifacts": persistent, structured interfaces matched to the domain. At Legora, these include an annotatable document editor (where users can highlight a specific clause and direct agent changes to only that clause) and a tabular review interface (a structured grid surfacing flagged items for human judgment). He closes with the observation that language is the universal interface for humans precisely because we lack other bandwidth — but agents are not humans and should not be constrained to human-scale communication primitives.