Agentic engineering for data analysis tools

Writing the spec first, then commissioning the agent to build inside it

Index of the documents in the ai-agentic-engineering-analysis project: an evergreen guide to specifying analysis work before an agent builds it, and the template new projects here start from.

Vibe coding accepts whatever the agent produces. Agentic engineering writes the requirements, the constraints and the checks first, then commissions the agent to build inside them. This project works out what that means for pharmacometrics and data analysis work, where the deliverable is usually an argument supported by numbers rather than a running service.

Two questions drive it. How is the initial specification written, and who writes it. Everything else follows from those.

The guide is evergreen: it is revised as the practice changes rather than finished. It is also the template that new projects in this repository start from.

Documents

• Working specification. What is already in place here, how to write the initial spec, how the work divides between Andy and the agent, what a new project starts with, and two published planning tools, Superpowers and Spec Kit.
• References. The sources this draws on, and a marker on each recording whether the claim drawn from it has been checked against the source.
• acop_abstract_2027_idea.md. A draft idea for an ACoP 2027 abstract on the pharmacometrics case: the five stages of an analysis, who reviews what, and what changes once an agent does the hands-on work. Not linked, because the site renders .qmd only and this one stays a working file in the repository.

Related reading

Context Engineering for AI-Assisted Pharmacometrics: A Practical Tutorial (Pritchard-Bell et al., 2026) describes task definitions that give an agent the domain rules, verification criteria and examples it needs at each step. Tasks start with fresh context and pass results through workspace files.

My reservation is the covariate-selection rule: the tutorial specifies an objective function value decrease of ΔOFV > 3.84 for inclusion and uses forward selection and backward elimination. I tend to prefer a full covariate model, with scientifically motivated covariates specified together, because stepwise selection can overfit noise. I would want the task definition to reflect that choice before asking an agent to develop the covariate model.