Sheva · Part 3 · AI Integrations

Now Build It.
Your First AI Agent.

From handing off one task — to setting up help that runs itself.

A self-guided guide · No code. No new tools. Just the next step.
CORE IDEA

An agent is a handoff that repeats

ACTIONABLE GOAL

Design one agent for your week

PERSONAL OUTPUT

Your own agent recipe card

Where We've Been

You already know how to do this.

Quick catch-up, whether or not you've done the first two guides.

1
You picked a tool

Copilot, ChatGPT, Claude, Gemini, Grok — or your company picked for you. Either way, you're set.

2
You learned to hand off

Spotting the repetitive, draining, blank-page tasks and letting AI take the first pass.

3
Now: you make it run itself

Take a handoff you already do — and set it up so you barely have to ask twice.

New to the series? You're still fine.
The whole series lives on my site — Part 1 and Part 2 are there. But this one stands on its own.
01 · The Big Idea

An agent is just a handoff that repeats.

Forget the sci-fi. Strip away the buzzword and an "agent" is something simple: instead of asking AI to do a task once, you set it up so it does that task again and again — with less and less input from you.

A prompt

You ask once. You get one answer. You start over next time.

A saved handoff

You save the instructions so you're not retyping them every time.

A multi-step handoff

One request triggers several steps — draft, check, format — in a row.

An agent

It knows your context, runs the steps, and only checks in when it needs you.

The through-line:
Every level is the same skill you already have — handing off. An agent just means you set it up once so you stop asking from scratch.
02 · What Every Agent Needs

Three ingredients. That's it.

Whether you build it in Copilot, a custom GPT, a Claude Project, or a Gemini Gem — every agent, at any level, comes down to these three.

A
A job — what it does, every time

"Turn my meeting notes into a summary and action items." One clear, repeatable task.

B
Context — what it always needs to know

Who you are, your format, your tone, your rules. Set once, so you never re-explain.

C
A trigger — when it runs

You paste something in, say a keyword, or (more advanced) it kicks off on a schedule.

Try this

Think of one handoff you already do. Say it out loud as a "job." Got it? You've done the hardest part.

03 · Where You Actually Build It

Three tiers. Pick your comfort level.

Same three ingredients, very different tools. Start wherever you are — you can always climb.

Tier 1 · No-code — start here

Built into tools you already have. No coding, ready in about ten minutes. Match it to your ecosystem.

Custom GPTs (ChatGPT) Claude Projects Gemini Gems (Google) Copilot agents (Microsoft)

Why it's the right start: you set the instructions, upload a few files, and it remembers your context every time. That's a real agent.

Tier 2 · Low-code / business — connect your systems

For agents that touch multiple apps and run real business workflows. A visual canvas, but a real learning curve.

Microsoft Copilot Studio Power Automate n8n (open-source)

Why here: these link an agent to your email, CRM, files, and databases — with governance controls. At work, only connect what your organization approves.

Tier 3 · Developer — full control

Code-first, for engineering teams building custom, production-grade agents.

OpenAI Agents SDK Codex Cursor Claude Agent SDK

Worth knowing: even here, the idea's the same — job, context, tools. You're just writing it instead of clicking it.

One idea worth knowing: portable "skills."
A new open standard (SKILL.md) lets you write an agent's instructions once and reuse them across many tools. The takeaway isn't the file — it's that the skill you build isn't locked to one product.
Try this

Which tier fits you today — not where you think you "should" be? Most people start at Tier 1. That's exactly right.

04 · "I'd Use That!" — Agent Recipes

Real agents, work and home.

The weekly-report agent
Paste your notesSummarizesYour formatReady to send
The meal-planner agent
What's in the fridge3 dinnersGrocery listSame every week
The inbox-triage agent
Paste an emailDrafts a replyYour voiceYou edit + send
The tutor agent (for the kids)
A homework topicExplains simplyGives an exampleQuizzes them
Try this

Which of these would you actually set up? Pick one — that's your candidate for the exercise below.

05 · Keep These In Your Hands

More power means more care.

An agent does more on its own — so the guardrails from Part 2 matter even more here.

Keep a human in the loop

The more steps it runs, the more you check the output before anything goes out.

Mind what it can touch

At work, only connect an agent to data and systems your organization approves.

Start supervised

Run it with you watching a few times before you trust it to run on its own.

Know how to stop it

Whatever you build, know how to pause or turn it off. You're always in control.

Rule of thumb:
Your intern got a promotion — not your job. More trust, earned gradually, always supervised.
06 · Your Turn

Design your first agent.

Fill in the three ingredients for one agent you'd actually use. Then hit Print to save it — it won't save on its own.

My agent recipe card
Try this

Read your recipe card back to yourself. Missing an ingredient? That's the gap to fill before you build. Bonus: text it to one person — saying it out loud makes it real.

07 · Make It Stick

Build one this week.

Not ten. One. Set it up, run it a few times supervised, and see how it feels.

This week, I'll build an agent that
so I stop doing
Try this

Tell one person the agent you're building this week. That's your check-in — have them ask you next week if it's running.

08 · Keep Going

You built something. Now grow it.

The one thing to remember:
You're not learning a new skill — you're leveling up the one you already have. A handoff you set up once is an agent.