WHERE WE BEGIN
We have different starting points.
Some students are writing code for the first time. Others already use R for research. Many have used a chat tool. A coding agent that works inside a project will be new for most of the room. We will use one common project, and each student can take steps that fit their current experience.
I need one step before we automate five.
The agent made a plot. I want to know what it did.
I can read the code. Let the agent do the repetitive part.
I trust my script. Show me why the agent helps.
I am here. Curiosity may arrive after the coffee.
These descriptions show approximate starting points. Your position can change from one task to another. The lesson gives the room a common language for asking what the tool should do and what the student needs to understand.
THE QUESTION FOR TODAY
What should we delegate, to which system, and with what evidence?
This question replaces a tour of AI brands. We choose the task before the tool, scale autonomy to the consequences of error, and decide what proof will be required before execution begins.
Adam Smith used the pin factory to show how specialization can raise productivity. AI creates a related coordination problem inside analytical work: a person frames the question, an agent proposes or executes steps, software records the result, and a person verifies and interprets the evidence. The benefit comes from both specialization and a clear handoff. Hayek adds that relevant knowledge is dispersed; Garicano shows how organizations route common and exceptional problems to people with different knowledge. Applied to Math Camp, the agent can handle bounded production, while the student supplies context, recognizes exceptions, and judges the evidence. Source: Adam Smith, The Wealth of Nations, Book I, Chapter I → Source: F. A. Hayek, “The Use of Knowledge in Society” → Source: Luis Garicano, “Hierarchies and the Organization of Knowledge in Production” →
Net value of delegation = time saved + quality gain −
specification cost − verification cost − expected error cost.
A faster agent does not automatically create a faster workflow.
Compare time saved and quality gained with the time required to
specify the task, verify the answer, and repair plausible errors.
If a WDI calculation saves twenty minutes but definitions,
revisions, years, and arithmetic require thirty minutes of
review, delegation has not reduced total analytical work. This
extends Coase’s transaction-cost insight: compare complete ways
of organizing work, not one isolated production step.
Source: R. H. Coase, “The Nature of the Firm” (1937) →
Linus Torvalds’s intervention in a 2026 Linux code-review discussion moves us past a simple argument about whether AI is useful. His operational test is whether the tool helps maintainers instead of merely creating more work for them. In policy analysis, ask the same questions: Does the agent reduce total work or transfer it to a reviewer? Does it surface errors while they can still be corrected? Can the accountable analyst inspect, accept, or reject the result? Source: Linus Torvalds, “Re: Linking Patchwork with Sashiko?” (2026) →
Meet the instructor, recognize the room's different starting points, and name what everyone should be able to do by the end.
Separate syntax recall from reading, testing, changing, and explaining consequential instructions.
Use the pin factory to ask who frames, executes, records, verifies, and interprets.
Distinguish models, chat, agents, harnesses, R, RStudio, Quarto, and Codex; then choose the tool after the task.
Step away from the screen and return at minute 60.
Examine the new bottleneck, cognitive debt, and the learner as a second product of the work.
Use verifiability and potential harm to decide what may be delegated and what must remain a human decision.
Specify goal, context, permission, and proof; then ask Codex to retrieve current official WDI data and prepare a time-pressured meeting brief while we inspect the work live.
Open the project, locate the readiness command, and begin the installation clinic at minute 120.
Yes. Coding now includes reading proposals, making precise changes, rerunning work, testing results, and preserving the reasoning. Prompting expands what we can attempt. Coding makes consequential instructions inspectable and repeatable.
MODEL, HARNESS, TOOL, HUMAN
Each part of the system has a specific job.
| Layer | Job | Decision left to the student |
|---|---|---|
| LLM | Generates language, code, or a tool request from context. | Whether the policy question matters. |
| Chat interface | Exchanges messages with a model. | Whether a response is grounded. |
| Coding agent | Reads, edits, runs, observes, and continues in a loop. | Whether its output should be trusted. |
| Harness | Assembles context, exposes tools, enforces permissions, and returns results. | Whether the available check tests the right claim. |
| R | Executes specified statistical and data operations. | Whether the chosen operation represents the intended concept. |
| RStudio | Keeps scripts, console, objects, files, and plots visible. | Whether the code is substantively correct. |
| Quarto | Combines prose, code, figures, and output in a reproducible document. | Whether the narrative follows from the evidence. |
Workshop analogy. The model is a language engine. The harness is the workshop around it: workbench, toolbox, clipboard, locked cabinets, and measuring instruments. The coding agent is the engine operating inside that workshop. The student remains the investigator who writes the work order, grants access, watches the work, and decides what to keep.
Text, code, or a tool request.
Context, tools, permissions, and feedback.
Reads, proposes, runs, and observes.
Frames the question and defines proof.
Capability asks whether a system can complete a task under specified conditions. Adoption asks whether people actually integrate it into a workflow. Impact asks whether its use changes productivity, quality, learning, or decisions. A benchmark can establish capability under benchmark conditions; it cannot, by itself, establish adoption or impact.
THE NEW BOTTLENECK
When generation becomes cheap, understanding becomes scarce.
Geoffrey Litt distinguishes two forms of understanding. We need to understand enough to verify: Does the code run, match the request, and pass the relevant checks? We also need to understand enough to participate: Why was this approach chosen, what assumptions does it contain, and what question should come next?
Herbert Simon framed the same bottleneck as an allocation problem: information abundance consumes attention. When an agent can produce many plausible analyses cheaply, the scarce resource becomes the analyst’s capacity to select, verify, explain, and integrate them. Source: Herbert Simon, “Designing Organizations for an Information-Rich World” (1971) →
Cognitive debt grows when an investigation changes faster than the people responsible for it can absorb. The code may run while nobody can explain why a filter exists, what a variable means, or how to change the analysis safely. In policy work, this debt weakens the analyst’s capacity to judge whom the result represents and what decision it can inform.
State what should happen and what evidence would count.
Inspect the diff, run a small example, and ask why.
Explain the purpose, assumption, expected effect, and one check.
Ask for a toy example or a quiz before accepting the change.
Before keeping substantial agent-written code, state its purpose, identify its input and output, name one assumption, predict one consequence of changing it, and run one check that could expose a mistake.
Today: qt = q(ht,
at). Tomorrow: ht+1 =
ht + practice + AI scaffolding − displaced practice.
AI assistance can improve today’s script, table, or explanation.
Its effect on tomorrow’s capability depends on whether it adds
scaffolding and deliberate practice or displaces the practice that
maintained the skill. A productive classroom workflow therefore
evaluates two outcomes: the artifact students produce now and the
work they will be able to understand, modify, and defend later.
Michael Polanyi argued that expert knowledge includes patterns, exceptions, and judgment that cannot always be reduced to a complete explicit rule. Workplace evidence from Brynjolfsson, Li, and Raymond found that AI assistance increased customer- support productivity by 14 percent on average and by 34 percent among novice and lower-skilled workers. One interpretation is that AI helped distribute patterns previously concentrated among experienced workers. The educational question remains whether students also learn when those patterns do not apply. Source: Michael Polanyi, The Tacit Dimension → Source: Brynjolfsson, Li, and Raymond, “Generative AI at Work” →
This emphasis on observable action and correction also has a longer history. Norbert Wiener’s account of cybernetics centers communication and feedback: a system can correct only what it can observe. Our classroom protocol turns that systems insight into a practical habit—watch the trace, compare the result with a prediction, and use the discrepancy to revise the next action. Source: Norbert Wiener, Cybernetics (1948) →
SOURCES Adapted and paraphrased from Geoffrey Litt, “Understanding is the New Bottleneck” (2026), and Margaret-Anne Storey, “How Generative and Agentic AI Shift Concern from Technical Debt to Cognitive Debt” (2026).
In Plato’s Phaedrus, writing can preserve words while giving a learner only the appearance of wisdom. Aristotle’s Nicomachean Ethics distinguishes technical making from practical wisdom: deliberating well about situations that could be otherwise. The Math Camp translation is concrete. An agent may extend memory and technique, but students still need to question and reconstruct what it produced and decide what should be done, for whom, and under which uncertainty. Source: Plato, Phaedrus 274c–275b → Source: Aristotle, Nicomachean Ethics, Book VI →
BEFORE EXECUTION
Write four things down.
Use the output to generate questions and verify them before drawing conclusions.
Keep consequential judgment outside the agent.
Automate with a visible check.
Require review, assertions, and permission boundaries.
Goal: Prepare an evidence inventory for this question: How is GDP per capita associated with under-five mortality across countries in 2022? Context: Use the shared frozen WDI data and indicator dictionary. Focus on country, year, gdp_per_capita_ppp, and under5_mortality. Permission: Read and summarize only. Do not edit or update the data, retrieve new data, or fit a model. Constraints: Use association language, flag missing values, and do not invent definitions. Proof: Report the unit, variable definitions, 2022 coverage, and three checks the class should complete before analysis.
This work order asks for an evidence inventory, not a conclusion. A good work order exposes failure, supports supervision, and gives the class a concrete result to verify before fitting a model or making a claim.
THE FIRST LAB
Analysis begins after every machine passes the readiness gate.
Lab 1 is an installation clinic. Every student must open the complete course project, start R, load the required packages, find Quarto, open the frozen data, and confirm that Codex can read the project.
Rscript code/check_setup.R --codex-confirmedGreen means ready. Yellow means the automatic checks pass and Codex confirmation remains. Red means repair before analysis. Every unresolved failure leaves with a named next step and a person responsible for follow-up.
Open the complete installation clinic →
Students who reach green while repairs continue complete an optional 8–10 minute read-only proof: ask Codex to locate the start instructions, frozen WDI data, and indicator dictionary; verify every reported path manually; and distinguish file evidence from the agent’s inference. This is the first supervised agent exercise, not a submission.
SLIDES
Lesson deck.
The slide deck is a 16:9 LaTeX Beamer PDF. It follows the same lesson and ends with the installation clinic.
The first teaching block contains three checkpoints and the second contains two. Every checkpoint reserves two minutes for discussion with nearby classmates and two minutes for whole-room sharing. Block 2 then ends with a live online policy-briefing demonstration. These are opportunities to test an idea and hear another interpretation; there is no submission or grade.