AI
Planning and Reviewing Are the New Bottleneck
Six to twelve months ago, completing complex work meant focusing on doing the work. That was the bottleneck. Today, models handle execution cheaply, and the constraint has shifted to planning and reviewing. Getting the specs right, capturing non-functional requirements, and then checking the output takes the time now. Reviewing a large pull request on GitHub is painful. Letting AI agents review their own work without humans in the loop might work, but often doesn’t.
The Verifier’s Rule: Where AI Can Actually Succeed
Jacob Lauritzen frames agent capability around the verifier’s rule: if a task is solvable and easy to verify, AI will solve it. Run the agent in a loop, tell it what it got wrong, and it converges. Legal shows the range clearly. Checking that a contract’s definitions are used consistently is easy to verify and easy to automate. Writing the contract is trivial to produce but almost impossible to verify, because the only real test is a court ruling. Litigation strategy is harder still: five lawyers give five answers, no objective truth, and AI makes little headway.
Building Agent Trust Through Decomposition and Guardrails
Three moves increase trust. First, use proxy verification: for contracts, test whether the new draft resembles known good contracts. That’s not ground truth, but it gives the agent feedback to improve against. Second, decompose the task. A full contract breaks into formatting checks, definition linting, and clause review. Leave the risk profile and negotiation stance to the human; automate the parts that have clear pass-fail answers. Third, add guardrails. Limit the agent to specific files and specific directories. Claude Code does this at the low-trust end, asking permission before every action, and forgoes it entirely in YOLO mode.
Control Requires Skills and Elicitation, Not Just Planning
Planning gives one upfront steering moment. After that, as Lauritzen described it, you’re working with a co-worker who aligns on the approach and then
“you never ever hear from them again” (Jacob Lauritzen)
Skills encode judgment at each node: whenever the agent reviews a termination clause, the relevant EU law is already in the skill. For gaps skills don’t cover, elicitation works. The agent makes a decision, logs it, and keeps going. The human reviews the log and reverses anything wrong. Legora’s tabular contract review (12:18) implements this: the agent flags uncertain items before resuming.
High-Bandwidth Artifacts Over Chat
Chat is one-dimensional. It collapses a tree of agent work into a linear stream, and if that tree grows 10 or 100 times, the chat becomes an endless scroll of 50 unanswerable questions with no context attached. The right interface is a persistent, high-bandwidth artifact: a document where you highlight clause three and only clause three changes; a tabular review where the agent’s flags are scannable at a glance. Language is a universal input method for humans, but
“but agents aren’t humans.” (Jacob Lauritzen)
Constraining agents to a chat box squanders that difference.
Notable Quotes
collaborate in high bandwidth artifacts. Jacob Lauritzen · ▶ 11:44
you never ever hear from them again Jacob Lauritzen · ▶ 9:35
but agents aren’t humans. Jacob Lauritzen · ▶ 13:31
Key Takeaways
- Doing work is now cheap; planning and reviewing are the bottleneck for complex agent tasks.
- If a task has no objective verifier, even strong AI agents will struggle to reliably complete it.
- Chat is a low-bandwidth interface; agents need persistent, structured artifacts for effective human collaboration.