How to use this site

Work top to bottom. Concepts first so the tools make sense, then hands-on practice, then you'll shape a project of your own. Check off boxes as you go — your progress saves in this browser.

What is "agentic" AI?

You've probably used a chatbot: you ask, it answers. An agent goes a step further — it can take actions in a loop to accomplish a goal you give it.

The agent loop

Almost every agentic tool works the same way under the hood:

   ┌─────────────────────────────────────────────┐
   │  1. You give a goal                           │
   │  2. The model decides on an action (a "tool") │
   │  3. The tool runs (read a file, run code...)  │
   │  4. The model reads the result                │
   │  5. Done? → answer.  Not done? → back to 2.   │
   └─────────────────────────────────────────────┘

What agents are good (and not good) at

The four words you'll hear all summer

LLM, agent, skill, MCP. These come up constantly in OPUS-AAI Week 1. Here's what each one actually means.

Set up Claude Code

Claude Code is a terminal-based agent. Don't worry if the terminal is new — you'll mostly type plain English to it.

# check if you already have node
node --version

# install Claude Code (once Node is present)
npm install -g @anthropic-ai/claude-code

cd ~/my-first-project
claude

• Terminal opens and I can type commandsTry pwd and ls
• node --version prints a version ≥ 18
• Claude Code installed and claude launches
• Connected to the program endpoint (mentor confirmed)

Your first session

The whole interface is a conversation. Type what you want in plain English. Here's a gentle on-ramp.

Try these, in order

> What files are in this folder, and what does each one do?

> Make a file called hello.py that prints today's date.

> Run it. If it errors, fix it and run it again.

> Now add a function that counts lines in a text file,
  and write a tiny test for it.

Notice what's happening: Claude Code proposes actions, you approve them, it reads the results and keeps going. That's the agent loop from Module 1, live.

Three things to know early

• Had a back-and-forth conversation with the agent
• Watched it create, run, and fix a file
• Approved (and declined) at least one proposed action

Starter exercises

Small, realistic tasks that build the muscles you'll use on a real project. Do at least three. Marine-data flavored, but no biology background required.

• Exercise A — tidy a CSV
• Exercise B — explain code
• Exercise C — plot data
• Exercise D — version control
• Exercise E — write a skill (stretch)

Working well with agents

The difference between frustrating and magical is mostly how you set things up. A few habits go a long way.

About the lab

You'll spend the summer applying these tools to real questions. The Roberts Lab studies how marine invertebrates — especially shellfish like oysters — respond to a changing ocean, using genomics and reproducible computation.

Project idea starters

Good OPUS-AAI projects pair a real lab need with the agentic methods you just learned — and produce a playbook others can reuse. Here are seeds. Make them yours, or invent your own.

Sizing your idea

A good 10-week project is narrow enough to finish and produce a clear deliverable. Run any idea through these:

Draft your pitch

Before you meet your mentor, fill in this one-paragraph pitch. It forces clarity and gives them something concrete to react to. (Practice for the Week 9 non-expert pitch and the Week 10 final presentation.)

PROJECT:   <one-line name>

PROBLEM:   The lab currently struggles to ______,
           which matters because ______.

APPROACH:  I'll use agentic AI to ______
           (which agent / skills / MCP tools: ______).

DATA:      I'll use ______ (already available? yes/no).

WEEK-2 GOAL:  A first working version that ______.

DELIVERABLE: A ______ plus a playbook documenting
             how others can reuse this approach.

UNKNOWNS:  The biggest thing I'm unsure about is ______.

• Skimmed the lab's notebooks / Current Findings / a repo
• Picked or invented a project idea
• Ran it through the sizing questions
• Filled in the pitch template
• Pressure-tested it with the agent
• Ready to discuss with my mentor

The playbook (your big deliverable)

Every team delivers a playbook: a practical guide to using agentic AI in your research area. The good news — if you take notes as you go, it nearly writes itself.

Capture as you work

Keep a running playbook.md from week one. Each time something works (or fails instructively), jot it down. By Week 10 you'll have a real document, not a panicked weekend.

A solid playbook usually includes

• The problem — what research task this addresses.
• Setup — tools, endpoints, and how to get running.
• Workflows — the step-by-step approaches that worked, with example prompts.
• Skills — any reusable skills you built, and how to use them.
• Gotchas — what went wrong, what to watch for, how to verify outputs.
• Results — what you actually accomplished, with a figure or two.
• Next steps — what you'd do with more time.

Resources & quick reference

Glossary

You're ready. Bring your pitch to your mentor, keep notes for the playbook, and remember: you're the scientist — the agent just helps you move faster. 🌊