Intro to AI coding assistants

Programming with LLMs for Data Practitioners

Julia Silge | Posit PBC

Hello!

@juliasilge

@juliasilge.com

juliasilge.com

Day 2: Build with AI

1. Intro to AI coding assistants (this session!)
2. Working in your editor
3. Steering and extending
4. Agentic workflows

Where we left off

• Yesterday we ended by sitting with the tension of real promise and also real harms
• Today we go deeper into the most powerful (and most fraught!) way to work
• The plan: let’s stay honest and stay in control

From “build on” to “build with”

• Day 1: you wrote code that calls LLMs (ellmer / chatlas)
• Day 2: AI assistants help you write and run code, on your own data
• Same models, a new way of working

A ladder of context

How most of us start

• Web chat: paste text into ChatGPT and copy paste back
• Chat in your editor: assistant can see your files, session, and errors
• Inline completions: assistant suggests code right where you are typing
• Agents: fetch their own context and act in a loop
• Going up the ladder, the assistant sees more of your world

When to reach for which

• Web chat: quick, throwaway questions (or maybe never?)
• Editor chat: questions and edits grounded in your real code
• Completions: small and in-flow where you stay fully in control
• Agents: multi-step work you would otherwise do by hand
• Two dials, not one
  • How much it sees
  • How much it does

What is an agent?

An LLM in a harness

• We built up to this yesterday by learning about turns, conversations, tools
• Harness: the program that wraps the model and runs its tools
  • a web chat, an assistant in your editor, a command-line tool
• Agent: an LLM in a harness that calls tools in a loop, deciding each step from the last result
• An agent needs a read tool (to observe) and a write tool (to act)

The question that follows us

• An agent is not useful unless it can take actions on your behalf
• But how do you know it is taking the right actions?
• We will keep returning to this all day: read vs. write, plans, tests, control

Meet Posit Assistant

One assistant, many surfaces

• A data-science assistant for Positron, RStudio, and the terminal
• Bring your own LLM
  • We’ll continue to use Anthropic in our workshop
• It knows about your R or Python session
  • For example, it can inspect your data frames and their schemas
• We’ll use Posit Assistant for three things
  • Explore your data
  • Build a dashboard
  • Debug and fix code

The top of the ladder, in one tool

• Editor chat, completions, and a full agentic loop, all in one place
• No more copy-pasting from a browser
• This is the tool we will use all day
• One assistant, one dataset, climbing from assisted to autonomous

Setting up Posit Assistant

• If you are using RStudio, you will need to sign up for Posit AI
  • Free trial for Posit AI should cover your workshop usage today
  • Read docs at https://posit-dev.github.io/assistant/docs/downloads/rstudio/
• If you are using Positron, you can use the workshop-provided Anthropic API key or sign up for Posit AI
  • Read docs at https://posit-dev.github.io/assistant/docs/downloads/positron/

Your turn

Activity

Get set up with Posit Assistant:

• Open Positron or RStudio and enable Posit Assistant
• Configure your model provider
• Confirm it responds: ask it to say hello

10:00

Cost and context

Remember tokens?

• Every turn replays the whole conversation and you pay per token
• Cached input is cheap and the output tokens are the expensive part
• Agents have long conversations, so this matters even more
• /context shows your usage
• /compact summarizes to save tokens

Context is the key to good help

• The assistant works best with high-signal context
• It already sees your environment including data, plots, errors
• So let it fetch what it needs; do not paste everything
• Give it what matters, trim what does not

Demo

How do cached input vs. output tokens drive the cost of a long, agentic conversation?

Walk through Simon’s interactive cost explainer

First contact

Our dataset: Utah crashes

• About 250,000 crashes, 2016 to 2019, from the Utah Open Data Portal
• Severity, location, and risk factors
  • DUI, distraction, night, teen driver, and more
• We will explore it, build on it, and (carefully) let an agent loose on it

Your turn

Activity

Start a chat with Posit Assistant and let it do the loading:

• Ask: “Read data/utah-crash-data-2020.parquet and tell me what is in it. What are the columns, and what might be interesting to explore?”
• Notice that it loads and inspects the data itself and you do not need to paste anything
• Try a follow-up question about something it mentions

10:00

The rest of today

• Working in your editor:
  • Completions, chat, explore and debug
• Steering and extending:
  • Plans, memory, skills, and MCP
• Agentic workflows:
  • More autonomy, with guardrails and reproducible output
• From assisted to autonomous, one crash dataset the whole way

Thanks!

@juliasilge

@juliasilge.com

juliasilge.com