Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Setup

Instructions to set up your environment for the workshop.

Workshop for today

We’ll be running in the same JupyterLab instance we’ve been using. Since these tools are CLIs, we can run them inside JupyterLab’s terminal.

We’ll be running OpenCode and the Nemotron 3 Super 120B model. You’ll need to set up the harness. We’ve provided a setup file at /u0b/software/training/opencode.jsonc that you can copy to ~/.config/opencode/opencode.jsonc. The contents of the file are:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "sdcc": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "SDCC",
      "options": {
        "baseURL": "https://inference0-api.sdcc.bnl.gov/v1",
        "apiKey": "{env:OPENAI_API_KEY}"
      },
      "models": {
        "nemotron-3-super-120b": {
          "name": "nemotron-3-super-120b",
          "limit": {
            "context": 131072,
            "output": 32768
          }
        }
      }
    }
  }
}

You’ll need to set OPENAI_API_KEY to the key you were provided for the workshop. You can also replace the {env:OPENAI_API_KEY} string with the contents of the key.

Code

This repository is https://github.com/henryiii/agentic-ai-workshop, which holds the workshop material only. You want https://github.com/henryiii/agentic-ai-example, which is the example project we’ll be working on. It was created with this prompt in Claude Fable 5:

Input

This is an example repo for a workshop. I want to start by implementing a library: let’s make a terminal-based plot tool, fairly similar to matplotlib, but for the terminal. We should make components for axis (doing tick marks is the hard part of plotting libraries, and we are more restrictive than normal being in a terminal with fixed chars!). We can do three plots for now: a standard line plot, a bar plot, and a 2d “imshow” style plot, using terminal colors. We can use dual colors per bin on that one; terminals have a half-height block and we can use bg/fg colors. A few guidelines:

  • We can make a demo app with textual (similar to https://github.com/scikit-hep/uproot-browser but of course minimal with sample plots and a couple of controls to change the plots) where students can see the plots. This isn’t part of the core library, but an example/app.py and textual can be in our dependency-groups in examples -> dev.

  • This is a demo repo, so don’t worry about fine edge cases, simple code is better than perfectly correct code under all circumstances. If it adds lots of code to handle NaNs or something, go with the simpler code.

  • Try to use high level OO design, don’t worry about exact API matching with anything (Matplotlib’s OO API if you do need a target, but simpler is fine)

Output

The model reported:

Done — the terminal plotting library is implemented, tested, and lint-clean. All 35 tests pass, prek (ruff, mypy, codespell, etc.) is clean, and pylint scores 10/10.

What was built, as flat modules in src/agentic_ai_example/:

  • canvas.py — Canvas, a grid of Cells (char + optional fg/bg RGB). put() silently clips out-of-range writes so artists need no bounds checks. render() emits ANSI truecolor codes only when the color changes, with resets so color never leaks across lines.

  • ticks.py — the teaching centerpiece: nice_ticks() (MaxNLocator-style snap to 1/2/2.5/5/10 × power of ten, snapping up so the count never exceeds what fits) and format_ticks() (one shared format per axis so label widths are predictable).

  • colormap.py — Colormap (linear interpolation over RGB control points) and a 5-point VIRIDIS.

  • artists.py — Transform (data → cell coordinates, y flipped), plus Line (● markers connected cell-by-cell), Bars (full blocks with ▁▂▃▄▅▆▇ partial tops), and HeatMap (two pixels per cell via ▀ with fg/bg colors).

  • figure.py — Axes (collects artists, unions their bounds, computes the y-label gutter after picking ticks so labels always fit, greedy-drops overlapping x labels) and Figure (print(fig) works via __str__).

The placeholder example.py is deleted, and __init__.py exports Figure, Axes, Colormap, VIRIDIS. The library is stdlib-only.

Demo app: examples/app.py is a ~80-line Textual app (textual lives in a new examples dependency group, included in dev). Keys 1/2/3 switch line/bar/heatmap, +/− adjust a parameter, q quits; plots resize with the terminal. I verified it headlessly with Textual’s test pilot. Run it with uv run --group examples python examples/app.py.

Two small deviations from the plan worth noting: I allowlisted bar in the pylint config (it’s in pylint’s default placeholder-name blocklist, but it’s the correct API name), and one test assertion was wrong on first run (expected a “3” y-label where nice_ticks(0, 3, 3) correctly yields [0, 2]) — the render itself was right.

I’ve left the changes uncommitted; let me know if you’d like a commit.

Followup

This is mostly the output (including a couple of bugs you’ll be finding), but I did follow up with “Is there a good way to make the textual example app more friendly to running headless, so a harness can control it and look at the output?” and guided it into making a better headless setup to run your AI harness on. Claude Code was happy enough without it, but this helps the smaller models with something they can run without getting stuck waiting for keypresses. This mostly added examples/headless.py and tests/test_app.py.

Outline for general use

If you are not part of the workshop, here’s an outline of the general setup steps.

Agentic AI Workshop
Agentic AI Workshop
Slides
Intro to Agentic AI