The Build Log · 5 min read · Oct 21, 2026
Why every AI agent I run gets a written permission lane
Agents are fast. That is the problem. A lane says who, what, how often, and where it gets logged.
My agents can send email, post to social accounts and update a calendar. Speed was never the issue. The issue was knowing, in writing, which of those things they are allowed to do on their own.
Early on I ran it on feel. An agent would draft something, I would glance at it, and it would go out. That works until the day you are busy, the glance gets skipped, and a message lands in the wrong inbox. One of those is enough to make the point.
What a lane is
A lane is a short, written permission for one kind of outward-facing action. Anything that touches a real person, such as an email, a text, a post or a reply, needs a lane. Anything outside a lane waits for me. Every lane answers four questions:
Who. The exact audience. "Customers writing in about their own account" is a lane. "Anyone who might be interested" is not.
Which channel. Email from one named address, or replies on one named account. One lane covers one channel.
How often. A daily cap and a follow-up rule. Most of my lanes allow one follow-up at most, and a "no" ends it.
Where it is logged. Every send gets a line in the day's log, with what was sent and why. If it is not logged, it did not happen.
The three questions before I grant one
First, would I be comfortable if this message were printed with my name under it? If not, it is not a lane. It is a draft for me to review.
Second, can the agent check its own facts before sending? Lanes work for things like account status, a public record or a published price, where the agent can verify against a live source. They do not work for judgment calls about price, terms or anything legal. Those stay with me.
Third, is there a cheap way to stop it? Every lane has a cap and a kill switch. If something looks off, I can close the lane in one line and nothing else breaks.
Lanes I closed
I have closed lanes too. The pattern is almost always the same: the audience was described too loosely, so the agent's idea of "who" drifted. The fix is never "be more careful." The fix is a narrower sentence in the lane and a test that proves the agent ignores everyone outside it, including its own earlier messages.
The result is less exciting than "fully autonomous AI." It is better. The agents do the routine work all day, and I only see the things that actually need me.
— Forrest
Get the next one in your inbox
One email a week: what ran, what broke, and what I changed.