The mistake almost everyone makes is starting too big. A great first agent does one small job well. Resist the urge to build an everything-assistant on day one. Follow these four steps and you will have something worth running.
Step 1: Pick one job
Choose a single task from Lesson 2 that you actually want off your plate this week. One job, one clear finish line. "Send me a 5-bullet morning news digest" is a first agent. "Run my life" is not.
Step 2: Write the instructions
Instructions are just plain writing, the same way you would brief a new assistant. Cover four things: what the goal is, how you want it done, what you never want it to do, and when it should stop and ask you.
Example
A real instruction block
"Every morning at 7am, read these three sources and send me the five most important new items as short bullets. Keep it neutral and skip anything older than a day. Never include opinion pieces. If a source is down, tell me instead of guessing."
Step 3: Set the guardrails
Guardrails are the limits that let you trust the agent. Decide them now, not after something goes wrong.
- What is it allowed to do on its own, and what needs your approval first?
- What is the spending limit, if it can spend anything at all?
- Which tools and accounts can it touch, and which are off-limits?
- How will it reach you, and how often is too often?
Step 4: Choose a model and define success
For a first agent, start with the default model your platform offers. It is almost certainly good enough, and switching later is easy. More important is defining success in one sentence, so you can tell whether the agent is actually helping.
That is a complete design: one job, clear instructions, sensible guardrails, a model, and a definition of success. It fits on an index card, and it is more than enough to launch. In the next lesson we connect your agent to the outside world, which is where guardrails earn their keep.