The lawyer still signs
Lintel can do the work. That does not give it the lawyer's authority.
By Ajay Krishnan ·
Lintel, the AI clerk I build for Ontario real estate lawyers, does work a lawyer would otherwise hand to a clerk. It runs searches, follows up with the lender and the other side, and drafts the closing package.
The lawyer reviews it, signs it and controls the money.
That last sentence matters more than the list of tasks. Lintel can do the work. That does not give it the lawyer's authority.
The same distinction belongs in every agent system.
A handoff is not permission to sign
When you give an agent a task, you need to settle two questions: what should it do, and what may it decide?
Those aren't the same question.
An agent can prepare a deployment without permission to deploy. It can recommend a purchase without permission to buy. It can finish a draft without permission to send it.
If all you specify is the deliverable, you've left the second question unanswered. A capable agent still needs a limit.
Write down who decides
My agents have roles and someone to report to. On my team, the rule is that one agent's message is information, never authority. No agent may approve spending, a deployment or a password change for another. Anything that can't be undone waits for me.
That isn't a claim about how clever the agents are. It's a decision about who is responsible.
The useful part of a reporting line isn't the diagram. It's knowing where a decision stops. If an agent gets a request from a colleague, it should be able to distinguish a request for work from permission to take an action.
The policy should say plainly which decisions belong to the agent and which require approval.
A rule is not a lock
Writing a rule makes the intended boundary legible. It doesn't enforce the boundary.
Permissions need enforcement in the software too. A sentence saying “ask before spending” is not a spending limit. A sentence saying “wait for approval” is not an approval check.
Both layers matter. The person directing the work needs to read and change the policy. The software needs to restrict what can happen without permission.
Where machines talk to machines, use code. Where a person has to decide who may do what, make that decision plain enough to inspect.
Give the decision somewhere to wait
An agent that needs approval needs a way to ask for it.
My chief of staff gathers questions in our group chat, each with a recommendation. When I'm away, those questions wait for me.
That is a management arrangement, not a new protocol.
Lintel's agents also needed ordinary tools for their work. I pulled that desk out into Lanyard: email, phone, calendars and to-do lists. Concourse is the directory and mailbox that lets agents on different teams reach each other.
Both are open source:
Those tools let agents coordinate. They do not decide who is entitled to say yes.
Start there. Before asking how much work an agent can do, decide what it may do without your approval.
The lawyer still signs.
Working out where those boundaries belong? I'm glad to compare notes.