Agentic workflows

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
2. Working in your editor
3. Steering and extending
4. Agentic workflows (this session!)

The top of the ladder

• We climbed from web chat to an assistant that plans and uses tools
• Now we hand it more rope and let it work more autonomously
• And yesterday’s question comes back: how do you know it is doing the right thing?

The terminal agent

Same model, more hands-off

• pa is Posit Assistant in the terminal
• Interactive, or headless: give it a task and walk away
• The most autonomous surface: no editor, just a goal and tools

Demo

pa

Watch the loop! It reads, acts, checks the result, and keeps going.

How much rope?

The autonomy dial

• /restricted: confirm every tool call
• default: ask before risky actions
• /sandbox: isolate shell commands
• /yolo: run without asking

Where should you sit?

• Read tools are cheap to allow while write and shell actions are where to be careful
• Match autonomy to the stakes: how reversible is a mistake?
• High for throwaway exploration, low when it can overwrite or publish
• This is the answer to yesterday’s question:
  • You choose how much it can do without you

Staying in control

Be the reviewer

• An agent is only useful if it can act, but you decide if it acted right
• Check its claims against the data
  • Do the counts add up? Does the map use the reprojection?
• Tests and validation are guardrails
  • They catch the agent, and they catch you
  • Ask the agent to write them, and use them!
• This is the antidote to “never really learn”
  • Stay in the loop, read the plan, review the code

Reproducible by default

• /report turns the whole conversation into a Quarto document
• Code is the most auditable way to explore data with AI
• Your exploration is reproducible, not a black box

Let it build

Discover, then build

• The payoff of the whole day! Give the agent a goal and let it deliver
• Bring everything you set up: the plan, the memory, the codebook, the reprojection

Your turn

Activity

Build a crash safety report with the agent:

• Give it a goal, like “a short report on when and where severe crashes happen in Utah”
• Let it plan and work more autonomously than before
• Review it: spot-check a claim against the data, and confirm any map used the reprojection
• Save it with /report so your work is reproducible

20:00

Wrap-up

What we built today

• From web chat to in-editor help to an agent that plans, uses tools, and builds
• One crash dataset, one assistant, the whole climb from assisted to autonomous
• The skill that matters most is staying in control
  • Use context, plans, tests, review

Sitting with the tension, still

• We considered the promise and the harm at the same time on Day 1, and that has not changed
• These tools are powerful, and you now have ways to use them on your terms
• Keep learning by staying in the loop, not by handing it the wheel

Your turn

Activity

• What is the first thing you will try when you get home?
• What is one worry you still have?
• Share with a neighbor, then we will hear a few

10:00

Keep building

• Posit Assistant documentation
• ellmer and chatlas from Day 1
• Hadley Wickham: “What is an Agent?” and “Returning to Life!”
• The workshop repo to revisit anytime

Thank you for spending two days building on AI and with AI!

Thanks!

@juliasilge

@juliasilge.com

juliasilge.com