AGENTR
AGENTR
Agentic Data Science with RAgentic workflows
Course description
AGENTR is the capstone of the MHDSR pathway: a practical, hands-on introduction to agentic data science, using large language models and coding agents to build, run, and reproduce data analysis pipelines in R from the command line. It builds directly on the R foundations of the earlier courses, and a working knowledge of R is required to get full value from it.
The course works through the full agentic workflow: choosing between frontier and local models, prompting and context engineering, calling R and shell tools from an agent via structured tool use, grounding output with retrieval over data and documents, packaging reusable skills and multi-step agent workflows, and evaluating agent output with tests, guardrails, and human review.
The emphasis throughout is on keeping the analyst in control: verification, provenance, cost, privacy, and honest reporting of AI-assisted work.
Learning objectives
By the end of AGENTR, participants will be able to:
- describe what agentic systems are, and where they help or fail in data science;
- set up a CLI-based agentic workflow for R projects (API keys, local runtimes, project scaffolding);
- compare frontier and local LLMs on capability, cost, latency, and privacy;
- write effective prompts and manage context for data analysis tasks;
- call R functions and external tools from a model using structured tool use / function calling;
- build retrieval over datasets, codebases, and documents to ground model output;
- package reusable skills and multi-step agent workflows for recurring analyses;
- evaluate agent output with tests and human review, and add guardrails;
- manage reproducibility, provenance, cost, and data governance of AI-assisted analyses;
- deliver a reproducible, agent-assisted analysis pipeline with an audit trail.
Schedule
| No. | Topic |
|---|---|
| 1 | Introduction to agentic data science: LLMs, agents & the CLI workflow for R |
| 2 | Frontier vs local models: access, cost, latency & privacy |
| 3 | Prompting and context engineering for data analysis |
| 4 | Tool use and function calling: driving R and shell tools from an agent |
| 5 | Retrieval-augmented workflows over data, code & literature |
| 6 | Agent skills and reusable multi-step pipelines from the command line |
| 7 | Evaluation, guardrails, provenance & reproducibility of agent-assisted work |
| 8 | End-to-end agentic analysis project |
Practical information
- Audience
- R users, data analysts, researchers, and PhD students who want to use LLMs and coding agents for real analysis work
- Prerequisites
- Working knowledge of R is required, plus comfort with the command line; the R foundations from BIOSTATR, ML4HOR, or METAR are the assumed baseline
- Format
- Live sessions, slides, notebooks, datasets, exercises
- Schedule
- Mon/Wed/Fri 18h00–20h00, Sun 10h00–12h00 (Morocco time)
- Tools
- R, RStudio, Quarto, a terminal, an LLM CLI / coding agent, API access and/or a local model runtime
- Assessment
- Reproducible agent-assisted analysis pipeline with an audit trail