Tool: gives the LLM access to information by running R code
library(ellmer)chat <-chat_anthropic()#> Using model = "claude-sonnet-5".chat$chat("what's the time")#> I don't have access to real-time#> information, so I can't tell you#> the current time.
library(ellmer)chat <-chat_anthropic()#> Using model = "claude-sonnet-5".time <-tool(fun = Sys.time,description ="Get the time",name ="time")chat$register_tool(time)chat$chat("what's the time")#> ◯ [tool call] time()#> ● #> "2026-11-12 08:40:46"#> The current time is **8:40:46 AM**.
The LLM can’t run R. So who called time()? 🤔
Using the rapid learning skill
1. Establish Learning Goals
What do you want to learn, and why?
How tool calling works in ellmer
Where are you starting from?
I’ve used tools in ellmer before but that’s it
How would you like to learn this?
Running code
How much time do you have?
10 minutes
Using the rapid learning skill
1. Establish Learning Goals
Primary learning goal: by the end of 10 minutes, be able to explain what happens between asking a question and getting an answer, and where a tool call fits in.
Supporting objective: compare a conversation with and without a tool call and spot the difference
Using the rapid learning skill
2. Locate and Access Resources
You want to learn by running code, so our resource is an R session with ellmer loaded.
Using the rapid learning skill
3. Adopt and Execute Learning Activities
Activity 1: Compare a conversation with and without a tool call
Start a conversation and ask a simple question.
Register a tool, then ask a question that will need the tool to answer.
After each question, run chat$get_turns() to list everything sent back and forth. What extra steps does the tool call add?
Using the rapid learning skill
3. Adopt and Execute Learning Activities
Without a tool call…
library(ellmer)chat <-chat_anthropic()#> Using model = "claude-sonnet-5".chat$chat("Say hi!")#> Hi there! 👋#>#> How are you doing today? I'm happy to help with whatever you need...chat$get_turns()#> [[1]]#> <Turn: user>#> Say hi!#>#> [[2]]#> <Turn: assistant>#> Hi there! 👋#> ...
Using the rapid learning skill
3. Adopt and Execute Learning Activities
With a tool call…
time <-tool(fun = Sys.time, description ="Get the time", name ="time")chat$register_tool(time)chat$chat("what's the time")#> ◯ [tool call] time()#> ● #> "2026-11-12 08:43:13"#> The current time is **8:43:13 AM**.chat$get_turns()#> [[1]]#> <Turn: user>#> what's the time#>#> [[2]]#> <Turn: assistant>#> [tool request (toolu_01CW...)]: time()#>#> [[3]]#> <Turn: user>#> [tool result (toolu_01CW...)]: "2026-11-12 08:43:13"#>#> [[4]]#> <Turn: assistant>#> The current time is **8:43:13 AM**.
Using the rapid learning skill
3. Adopt and Execute Learning Activities
With a tool call…
#> [[1]]#> <Turn: user>#> what's the time#>#> [[2]]#> <Turn: assistant>#> [tool request]: time()#>#> [[3]]#> <Turn: user>#> [tool result]: "2026-11-12 08:43:13"#>#> [[4]]#> <Turn: assistant>#> The current time is **8:43:13 AM**.
What do you see in the second turn?
The assistant turn is calling time()?
Is it calling it, or requesting it?
Oh. Requesting.
Right. The model never calls anything. It only asks. And what do you see in the third turn?
Using the rapid learning skill
3. Adopt and Execute Learning Activities
With a tool call…
#> [[1]]#> <Turn: user>#> what's the time#>#> [[2]]#> <Turn: assistant>#> [tool request]: time()#>#> [[3]]#> <Turn: user>#> [tool result]: "2026-11-12 08:43:13"#>#> [[4]]#> <Turn: assistant>#> The current time is **8:43:13 AM**.
There’s a user turn with the result, but I didn’t type that.
So who ran time()?
…ellmer?
And who did it send the result to, and as whom?
To the LLM. As me?
Yes. Why would it need to do that?
So the LLM has what it needs to answer.
Using the rapid learning skill
4. Monitor and Evaluate Performance
Checkpoint before we move on. In your own words: who decides to use the tool, who runs it, and who sees the result?
The LLM decides, ellmer runs it, and the LLM gets the result as if I’d sent it.
That’s the whole mechanism, and you got there yourself. 🎉 Activity 1 done, five minutes under its timebox.
Using the rapid learning skill
5. Reassess Learning Strategies
What was most valuable about this session?
Comparing the two conversations side by side.
Noted. I’ll lead with a comparison next time.
What did this give me?
Huge reduction in overhead planning learning
Take on more ambitious tasks
More confidence in my learning
Use AI to customise learning to make it work for you
Learning through experimentation
Laboratory set from “Just Imagine”, 1930
“What can I work on that I can be sure is having an impact?”
March 2026 with Posit Assistant, using Claude Sonnet 4.6 and Claude Opus 4.5
A new metric…
Wait… whut?!
A poor experience for newcomers?
These PRs are sitting entirely unacknowledged, which is both a poor contributor experience and a growing triage burden for maintainers.
Some are substantive pieces of work (look at <list of PRs>) that just didn’t make it through.
😢
🤔
Don’t add judgemental comments about poor contributor experience please, just report the facts.
Hmm… how do you know? These were opened within 3 mins of each other…
A revelation
What happens to new contributors’ PRs?
New categorisation scheme
“There was a comment in the past 2 weeks”
New category: “active”
“Last comment was from maintainer over 2 months ago with no response from author”
New category: “abandoned”
“Last comment by the author over a month ago, still waiting for a maintainer reply”
New category: “stale-risk”
Back to my original question
What can I work on that I can be sure is having an impact?
Find stale PRs and either:
answer the question myself
ping the maintainers for an answer
Find abandoned PRs and either
finish them off
close them
What I learned
Wrong answers were still useful
Fast iteration made me realise I was asking the wrong question
Domain knowledge and scepticism did the hard work
Use AI to run experiments cheaply and learn about your reality
Adapting how you work
Alex Schomburg, cover for “The Exile of Time”, Avalon Books, 1964
Arrow in January 2026
Loads of new code contributions
Way too many tests
Never interact much
Arrow’s AI guidelines
“Only submit a PR if you can debug and own the changes yourself”
“Be upfront about AI usage”
The effect
nobody paid attention
frustrating experience reviewing code
arguing with other maintainers
Automatic AI code review with Copilot
Not so bad, right?
Things had changed more than we thought
Trying to keep up
all the caffeine!
just need to try harder
How do air traffic controllers maintain concentration?
Decision fatigue
Rethinking my approach
“How does <prominent maintainer> say no?”
Thank them
Explain why
Move on
A PR is a proposal, not a gift
A proposal
Not a gift
The actual solution
Set boundaries!
What does setting good boundaries look like?
You are allowed to say “no” without long justifications
I’m not going to read a 300-line AI-written PR description
What does setting good boundaries look like?
Focus on your own actions, not theirs
I can’t stop people submitting AI-generated pull requests, but I choose which ones get my time
What does setting good boundaries look like?
“If X then Y”
if(human_involved){engage()} else {close_pr()}
If there are signs a human is engaged, I’ll review their code
If not, I close it without one
AI Reviews
Copilot reviews the PR, I clean up after it 👎
AI drafts a review locally, I read it, own it, and act on it 👍
Cory Doctorow, via Wes McKinney at posit::conf
