THE NUMBER
Last fall, Roofing Contractor magazine asked its readers whether they planned to use AI, and only 9% said no, down from 32% the year before. Roofers, of all people, have mostly stopped asking whether to put the first one to work. The harder question is what happens once you have two.
3 THINGS HAPPENING RIGHT NOW
A Taco Bell franchisee asked a narrow question and got 20% more late-night sales at those stores
Charter Foods runs 466 restaurants, Taco Bells among them. One question senior director John Rankin put to an AI tool was one he could act on: should certain locations stay open later or close sooner? “The time to get that data gets significantly sped up, so we’re able to make that decision sooner,” he told Restaurant Business last fall. They opened those stores later, and late-night sales at them rose 20%.
A bankruptcy lawyer’s five agents spent $150 on their first day
Scott Bell, a bankruptcy lawyer, handed chunks of his forms-heavy work to five AI agents in March. They ran up $150 on day one, so he signed up for two $200-a-month subscriptions. He also had to teach them to ask less often, or the provider would throttle them. They now do several hours of repetitive work a week, from reading court notices to messaging clients, The New York Times Magazine reported.
THE DEEP DIVE
He stopped being the switchboard
Donnie McClanahan runs several restaurants. A broken espresso machine used to mean, in his words, “a phone call to the manager, who walks to the register, who scrolls through items one by one, who forgets one.” Every message he relayed could leak.
Now he asks his agent. When the machine went down at Tombras Cafe, he told it to mark every craft coffee out of stock and put them back in the morning. It reads sales, sees the mozzarella sticks running low, and asks, “Do you want me to set a reminder to have Jessica order more?” Then, unprompted, it went through his item catalog and flagged 11 problems, including two items with duplicate tax rates. “Nobody asked it to do that.”
His rule: “It proposes, I confirm.” He is part of Square’s AI champion program, so weigh the enthusiasm; the details are his own.
The real math: he handed over the relaying, meaning the call, the manager’s hunt and the item that got missed. He kept the one-word decision. Those are different jobs, and only the first needed him in the middle. He is now building a second agent.
Same split, any owner who relays messages between a system and staff all day: a dental group, a salon chain, a gym.
ONE THING TO TRY THIS WEEK
Week 2 of 4: Your second agent, and the send-back lane
The restaurant owner still gives every yes himself. Most yeses are routine, and routine checks are exactly what a second agent can take off you. Here is a doer, a checker, and a lane for sending bad work back, so you only see what fails twice.
Pick one job an agent or any tool already does for you that you look over before using, where a mistake that slips through is easy to fix. Now switch seats: imagine a stranger just handed you that finished job. Write the three lines you would want on top of it: what was done, what they weren’t sure about, what’s still missing. That is your handoff note.
Finish this sentence three times: “I’d send it back if…” For example: the total doesn’t match, a customer’s name is missing, it promises a date. These are your send-back reasons.
On a Mac open Terminal, on Windows open PowerShell, type
claudeand press enter. Paste: “Set up two agents for this job: [describe it]. The first does the job and ends every run with this note: [paste your three lines]. The second is a checker. It reads only the note and the finished work, not the first agent’s working, and checks them against these reasons: [paste your reasons]. If one applies, it sends the work back to the first agent with the reason in one line. If none applies, the work goes ahead. If the same job bounces twice, it comes to me. Anything touching money, legal, or anything that can’t be undone always comes to me. Keep a one-line log of every job, passed or sent back, set it all to run on its own, and tell me when it’s ready.” Claude builds both agents and the lane between them. Connecting them to your inbox or tools can take one setup step, and Claude will tell you what it needs.Let it run for a week. Each morning, skim the log: one line per job, marked passed or sent back. No lines means nothing ran. A rising count of send-backs means the first agent needs fixing. Once a week, look over one job the checker passed. If it let a bad one through, tell Claude to add that reason to the checker’s list. The handoff is a three-line note because, as Anthropic (the company that makes Claude) puts it, passing work between agents turns into a “telephone game,” with each handoff losing detail. A set format keeps the important parts. The checker never sees the first agent’s working, so it judges the result, not the reasoning behind it. Anthropic also says two agents can cost three to ten times as much to run as one, so add the second only when its job is genuinely different, like checking work it didn’t write.
Stuck? Reply to this email. I’ll help.
WHAT’S COMING
Next issue: what keeps two agents working the week something they depend on changes and nobody tells them.
Manu