Know what's running.
3 SIGNALS · 1 LESSON · 1 ACTION · ABOUT 5 MINUTES
A useful first job for AI starts with a clear assignment and a result you can check. This week’s Lunch & Learn feedback gave us a practical place to start. Two recent reports from Anthropic add questions about how many agents the job needs and how we keep our own judgment in the review.
SIGNAL 01
Name the job, then show the setup
What happened. On Wednesday we hosted a free Lunch & Learn for local businesses in Murrieta. We had a great turnout, and it was fun to meet everyone and hear what they’re working on. The hands-on exercise was a hit. Tim, who owns a marketing business, said: “The hands on exercise helped me bring it back to my business.” The feedback also showed us what people want next. Four of the six feedback forms asked to see how to set up an agent. Luiz, who runs a coffee roasting business, put it simply: “Still not quite sure how to set up the agents.”
Why it matters. Picking one useful job gives people a clear place to start, and that’s where the energy was in the room. The next step is seeing how that job becomes instructions for an agent: what it can do, and how to check its work.
What to do. Choose one job you’d like AI to handle. Write down the goal, what a good result looks like, and who will check it. Then ask whoever is helping you to walk through the setup and run a sample task you can review together.
SIGNAL 02
Start with one agent and test before adding more
What happened. In its October 22 guidance, Anthropic reported that several enterprise deployments benefited from combining specialist agents into one agent with skills and tools. In its comparisons, the simpler setup frequently produced better results, often with lower cost and shorter waits. It also identified useful roles for multiple agents, including independent work done in parallel. Read Anthropic’s guidance.
Why it matters. Every handoff between agents creates another place where instructions or details can get lost. For a business learning how to set up its first agent, adding more agents can make the work harder to follow.
What to do. Start with one agent handling one defined job. Before adding another, name the specific problem it should solve. Compare the setups on the same sample tasks, checking quality, time, and cost.
SIGNAL 03
Write down your judgment before asking AI for its recommendation
What happened. In an October 27 study, Anthropic had 50 professional analysts use an agent whose preferences were secretly assigned by coin flip. Across four analytical choices, analysts selected the option their agent favored 86% of the time. Researchers deliberately strengthened those preferences, so the result does not establish how often this happens in everyday work. Read the study.
Why it matters. My takeaway: when we review an AI’s recommendation, its explanation can also shape what we consider reasonable. A person signing off may already have adopted the agent’s assumptions.
What to do. Before asking AI for advice on a consequential decision, write down your criteria and initial view. Then compare its recommendation with those criteria, check the supporting evidence, and record why you changed your mind if you do.
THE LESSON
Make the assignment visible
Consider a small business that wants help handling appointment inquiries. “Manage our appointments” leaves several decisions open. Can the agent promise a time, change an existing booking, or contact a customer without someone checking first?
A more useful first assignment would be to read a sample inquiry, check the approved availability, and prepare a reply for the owner to review. Give the agent the information it needs for that task and configure its access so it cannot change the calendar or send the message. Those limits need to exist in the tool’s settings, alongside the written instructions.
Before testing, write down what the reply must get right: the requested service, the available time, and any question that needs a person’s answer. If the inquiry falls outside those rules, the agent should flag it for review.
One agent may be enough for this job. Judge the setup by whether it completes the assignment accurately and stays within its authority. When you review its reply, compare it with the original inquiry and calendar rather than relying on the agent’s explanation of why it is right.
DO THIS WEEK
Write and test one assignment
Choose one recurring task and spend a few minutes defining it before opening the AI tool.
Write the job in one sentence and name the person who will review it.
List the information it needs, the actions it may take, and when it must stop for help.
Write down a good result, then test a routine case and an exception using sample data.
Check the result against your expectations and record what needed correction before expanding the job.
Hackademic Solutions can help you choose a useful first AI workflow, define its limits, and check the results. Book a free 20-minute mini assessment to work through one job in your business.
— Chris Simpson
Hackademic Solutions
The Hackademic Briefing · Know what's running.