As agent production gets cheaper, planning and reviewing complex work become the main bottlenecks.
2
Agents work better when tasks are made verifiable through tests, proxies, decomposition, and guardrails.
3
Complex agent collaboration needs persistent, high-bandwidth artifacts instead of a single chat thread.
Summary
Jacob Lauritzen argues that complex agents need interfaces and workflows beyond chat. At Legora, a legal AI company, he sees the cost of doing work falling while planning and reviewing become harder. He uses verifier's rule to explain why agents perform well on tasks with clear checks, such as contract linting, and struggle with contracts or litigation strategy where correctness is difficult to establish. Teams can improve results by narrowing tasks, using golden documents as verification proxies, decomposing work, and limiting agent permissions. Lauritzen compares several ways for humans to guide agents. Up-front planning gives control but requires the human to understand the whole task. Skills encode judgment inside individual work steps and can handle cases discovered during execution. Elicitation lets the agent ask for help, while decision logs let it continue when nobody is available. His central product argument is that documents, tables, and other persistent artifacts give people more control and context than an infinitely long chat.
The hard part of agent work has moved from doing to planning and reviewing
Lauritzen says the economics of production have changed in the past 6 to 12 months. Completing the work itself is now often cheap, while planning and review take more time. Teams must define non-functional requirements and specifications before execution, then inspect the result afterward. He compares this to reviewing large GitHub pull requests, which he calls painful. Having agents review their own work may remove humans from the loop, but he is unsure whether that approach works reliably.
Agents solve tasks more reliably when correctness is easy to check
Lauritzen applies verifier's rule to agents: when a task is solvable and easy to verify, an agent can run in a loop, receive feedback, and fix its errors. Legal work shows the range. Checking contract definitions is easy to verify. Writing a contract is harder because its language may only be tested when a judge considers it in court. Litigation strategy is harder still because five lawyers may give different answers, with no objective answer to check. Coding has the same split, since some tasks are easy to test while building a successful consumer app is difficult to verify.
Humans should keep judgment where the work has no reliable test
The human-agent division should depend on how verifiable the task is. Lauritzen describes control as the ability to put human knowledge into the agent's work, and trust as how much of the trace a person needs to inspect. Humans can retain decisions such as a legal risk profile, precedent documents, and negotiation stance. Agents can handle work with clearer checks, such as formatting a contract or linting whether definitions are used and defined. This lets people spend attention where correctness cannot be established mechanically.
Verification proxies can make difficult work easier for agents
A task does not need a perfect test if it has a useful proxy. In coding, browser access and test-driven development can turn feature implementation into something more verifiable. For contracts, Legora can compare a new document with previous 'golden contracts' that are known to work well. The comparison does not prove that the new contract is correct, but it gives the agent a concrete signal. Decomposition adds more checks by breaking contract drafting into smaller operations with clearer expected behavior.
Planning gives control before execution but misses facts discovered during the work
Lauritzen models complex work as a tree or DAG of tasks. Without intermediate control, a person can only guide the agent at the root and then receives the finished result. Planning improves this by letting the person agree on clauses, review steps, and the general approach in advance. Its weakness is that the human must understand the entire task before execution. A special clause in one contract may only become visible during review, so the planning step cannot always provide the information needed to handle it.
Skills encode instructions at the nodes where work happens. For example, a skill can tell an agent how to review confidentiality or how to handle a special EU law when reviewing a termination clause. This supports contingencies discovered during execution and enables what Lauritzen calls progressive discovery. The agent can apply the relevant judgment when it reaches the case. Skills are more useful than a complete up-front plan for this reason, although no team will have a skill for every possible situation.
Elicitation and decision logs keep agents moving when they lack guidance
When a skill does not cover a situation, the agent can ask the human what to do. Lauritzen says the agent should avoid becoming blocked. It can make a decision, record the uncertainty in a decision log, and let the person review or reverse that decision later. This approach separates execution from immediate human availability. It also creates a record of the judgment calls that matter, rather than forcing the user to answer every question before the agent can begin.
Persistent artifacts give complex collaboration more bandwidth than chat
Lauritzen rejects chat as the main interface for large agent tasks. A large work tree collapsed into one linear thread would force a user to answer many questions without enough context. He prefers persistent, high-bandwidth artifacts that vary by industry. In legal work, a document lets someone highlight clause three, change only that clause, add comments, tag people or agents, and hand off sections to specialists. Legora's tabular review gives users a compact view of flagged contract issues, making agent output easier to inspect before continuing the remaining work.
"If you can make a task verifiable, you can just run an agent in a loop and tell it, "Hey, you did this wrong. Please fix it." and it'll eventually get there."03:40
Who should watch
You are building agents that run for a long time and need people to review or redirect them during execution.
Your team is deciding whether to use chat, planning prompts, skills, or a structured workspace for agent collaboration.
You work in legal, finance, coding, or another field where some tasks are easy to test and others depend on judgment.
