Presenters
Source
Beyond the Chatbox: Why Complex AI Agents Need More Than Just Conversation 🚀
If you have ever tasked an AI agent with a long-running, complex project—like drafting a legal contract—you have likely hit the “context rot” wall. You watch the agent spin up sub-agents, search the web, and write files, only for it to stumble on a minor detail, hallucinate, or lose the plot entirely 30 minutes later.
Jacob Lauritzen, CTO of Legora, argues that we are hitting a ceiling with the traditional chat interface. As we move toward more autonomous vertical AI, the way we collaborate with these systems must evolve.
💡 The New Economics of AI Production
In the last 6 to 12 months, the economics of AI have shifted dramatically. Doing the work is now extremely cheap, but planning and reviewing the work have become the new bottlenecks.
When we talk about complex agentic workflows, we often rely on the Verifier’s Rule: If a task is solvable and easy to verify, AI will solve it.
- The Challenge: Not all tasks are easily verifiable. While checking a definition in a contract is simple, determining a litigation strategy is nearly impossible to verify objectively.
- The Impact: When verification is difficult, humans must step in. The goal for vertical AI companies is to balance human control with agentic autonomy.
🛠️ Increasing Trust and Control
How do we make agents reliable enough to handle complex tasks without constant hand-holding? Jacob highlights two critical levers:
1. Increasing Trust 🤝
Trust is the inverse of how much you need to review an agent’s work. To boost it:
- Bring the task down the spectrum: Use techniques like Test-Driven Development (TDD) or browser access to make tasks more verifiable.
- Use Proxies: In legal, we can’t verify a contract until it hits a courtroom. Instead, we use golden contracts (precedent documents) as a proxy for success.
- Guardrails: Limit the agent’s scope—restrict the files it can edit or the websites it can visit. This prevents the “YOLO” mode that leads to accidental production database deletion.
2. Increasing Control 🎯
Control is the ability to instill your specific knowledge into the agent’s process.
- Planning vs. Skills: Planning is often inefficient; you essentially have to do the work to tell the agent how to do it. Skills are better—they allow you to encode human judgment into specific “nodes” of the workflow, handling contingencies (like specific EU laws) automatically as they arise.
- Progressive Discovery: If an agent hits a snag, it should make a decision to unblock itself while logging that decision for later human review.
🌐 Moving Beyond the Chat Interface
Jacob posits a bold truth: Chat is a low-bandwidth, one-dimensional interface. It attempts to collapse a complex, multi-branching tree of work (a DAG) into a single, linear string of text.
For complex work, we need high-bandwidth, persistent artifacts:
- Durable Interfaces: In legal, this looks like a document where humans and agents can collaborate via highlights, comments, and tag-based hand-offs.
- Tabular Reviews: By spinning up a table that summarizes the status of dozens of contracts, a human can spot issues in seconds rather than reading through 30 minutes of chat logs.
👾 The Future: Don’t Constrain AI to Human Limits
“Language is the universal interface for humans,” Jacob notes, “but agents are not humans.”
While chat is a great entry point, we shouldn’t limit our agents to the constraints of human conversation. We need to build systems that allow agents to interact with structured data, org charts, and persistent workspaces.
The takeaway? Stop trying to force complex, non-linear work into a single chat box. Build the workspace that the task requires, not the one that the chatbot demands.
Are you building the future of vertical AI? Legora is currently hiring engineers in London—reach out to Jacob if you want to join the journey! 🚀👨💻