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

Key Concepts

• Vertical AI: AI solutions tailored to specific industry workflows (e.g., legal, finance) rather than general-purpose models.
• Complex Agents: AI systems capable of multi-step reasoning, tool use, and long-running tasks.
• Verifier’s Rule: The principle that if a task is solvable and easy to verify, AI will eventually solve it; if it is difficult to verify, it remains a bottleneck.
• Context Rot: The degradation of an agent's performance or memory during long-running tasks due to context window limitations or "compaction" processes.
• High-Bandwidth Artifacts: Persistent, structured interfaces (documents, tables, canvases) that allow for more efficient human-AI collaboration than linear chat interfaces.
• Control vs. Trust: The trade-off between how much a human steers an agent (control) and how much they rely on the agent to act autonomously (trust).

1. The Evolution of AI Workflows

The speaker, Jacob (CTO of Legora), argues that the economics of AI production have shifted. Previously, the primary challenge was "doing the work." Today, because execution is cheap, the new bottlenecks are planning and reviewing.

• The Problem with Chat: Current agent interfaces often rely on linear, one-dimensional chat. This forces complex, multi-step tasks into a single thread, leading to "context rot," poor visibility into sub-agent actions, and difficulty in correcting specific errors (e.g., fixing one clause in a contract without breaking others).
• The Verifier’s Rule: Coined by Jason, this rule suggests that AI excels when tasks are verifiable. In legal work, "checking definitions" is easy to verify (and thus easy to automate), while "litigation strategy" is subjective and nearly impossible to verify, making it a poor candidate for full automation.

2. Strategies for Increasing Trust and Control

To manage complex agents, developers must balance human intervention with agent autonomy:

• Increasing Trust:
  • Task Decomposition: Breaking large tasks into smaller, verifiable sub-tasks.
  • Proxy Verification: When objective truth is unavailable (like in contract drafting), use "golden" precedents as a benchmark to verify if the agent’s output aligns with established standards.
  • Guardrails: Limiting the agent’s scope (e.g., restricting file access or directory read/write permissions) to prevent unintended actions.
• Increasing Control:
  • Planning: Aligning on an approach upfront. However, the speaker notes this is often inefficient because the agent lacks the context of the actual work until it begins.
  • Skills: Encoding human judgment into specific nodes of the agent’s workflow. Unlike planning, skills allow for "contingencies" and "progressive discovery," where the agent handles edge cases (like specific EU laws) automatically based on pre-defined logic.
  • Elicitation: When an agent hits an unknown, it should make a decision to stay unblocked, log that decision, and allow the human to review/reverse it later.

3. Moving Beyond the Chat Interface

The speaker posits that chat is a "low-bandwidth" interface that constrains agents to human-like communication. He advocates for High-Bandwidth Artifacts:

• Durable Interfaces: Instead of a chat box, agents should interact with persistent objects like documents or tables.
• Case Study (Legal): Legora uses a "tabular review" interface. Instead of a chat stream, the agent presents a table of contract clauses. The human can quickly scan, flag issues, and provide feedback, which the agent then incorporates. This allows for high control and rapid verification.
• The "Non-Human" Argument: Humans are limited by language because we cannot easily share complex mental models (like an org chart) without describing them. Agents are not limited by language; therefore, forcing them to communicate solely through text is an unnecessary constraint.

4. Synthesis and Conclusion

The main takeaway is that for vertical AI to succeed in complex domains, developers must move away from "chat-first" architectures.

• Actionable Insight: Build interfaces that mirror the way professionals actually work—using documents, spreadsheets, and visual workflows.
• Final Perspective: The future of AI collaboration lies in high-bandwidth, persistent artifacts where humans can inject judgment at critical junctures, while agents handle the heavy lifting of execution and verification within defined guardrails. As the speaker notes, "Agents aren't humans, and so we should not constrain them to human language."