Welcome to Day 3! Today, we’re going a level deeper: how they actually work behind the scenes — in human terms.

🧠 What You’ll Learn Today

• What “tools” really are (and why they matter)
• How agents remember things (and where they don’t)
• How agents create and follow a plan instead of just answering once

🧰 Tools: How Agents Touch the World

On their own, language models are just very good talkers. To actually do things, agents need tools. You can think of tools as the apps, services, and actions an agent is allowed to use.

Examples of tools an agent might have:

• Web: look things up, check sources, pull in fresh data
• Tasks and issues: read and create items on a to-do list or project backlog
• Files and repositories: read, create, and edit the files you're working on

You’re always in control of which tools an agent can access and what permissions it has.

🧠 Memory: What Agents Remember

If tools are how agents “touch” the world, memory is how they keep track of what’s going on.

There are two kinds of memory to know:

1. Short-term memory

   • What you’ve said in the current session
   • The steps the agent has already taken
   • The feedback you have given about certain choices or directions

2. Longer-term memory (in some tools)

   • Preferences you’ve shared (ex: “don't use em dashes”)
   • Important details (ex: your job title, your employer)
   • Specific rules to follow, commonly stored in a specific file or in your account preferences (ex: "always use camel case for Javascript")

Good agents remember what’s useful and appropriate for helping with tasks and many will ask or indicate when storing anything long-term.

You should always know:

• What it remembers
• Where that memory is stored
• How to clear or reset it

If you need help finding this info, just ask the Agent.

🧩 Plans: How Agents Break Work Into Steps

Finally, we get to the “agent-y” part: planning.

Instead of answering once and stopping, an agent will:

1. Understand your goal

• “Help me plan a holiday party for 15 people.”

2. Break it into steps

   • Pick a date
   • Draft an invite
   • Decide on food and drinks
   • Track RSVPs

3. Decide which tool to use for each step

   • Calendar for dates
   • Docs or email for the invite
   • Spreadsheet or list for RSVPs

4. Execute, check, and adjust

   • If something fails (e.g., a time slot is busy), try another
   • If something is unclear, ask you a follow-up question

The goal → plan → check → act → adjust loop is what makes agents feel more like a helper and less like a sounding board.

🛠️ Hands-On Experiment: Holiday Gift List 🎁

Let’s see tools, memory, and planning working together to create and manage a holiday gift list for a few people you care about.

Step 1 — Pick your agent tool

Use Gemini or any agent-enabled experience that can work with docs and a calendar.

Step 2 — Give it a clear goal

Try a prompt like:

You are an AI agent helping me with holiday gifting.

1. Ask me who I need to buy gifts for and a rough budget for each person.
2. Suggest 3 gift ideas per person.
3. Create a gift list in a document I can update.
4. Add a reminder on my calendar a week before I need to have everything ordered.

Step 3 — Watch for the three building blocks

As you go, notice:

• Tools: Does it ask to create a doc or add something to your calendar? Does it tell you which app it’s using?

• Memory: Does it remember people’s names and budgets as you answer? Does it carry details from one step to the next?

• Plans: Does it lay out a sequence of steps? Can it adjust if you say, “Actually, we lowered our budget,” or “Add one more person”?

📬 Keep the Streak Going

Next time: a demo of agents in action!

👉 Subscribe at 12DaysofAgents.com to get Day 3 automatically
👉 Forward Day 2 to a friend who would love an easier December 🎁