In 2023, language models autocompleted code. In 2026, agentic systems read whole repositories, run code and inspect the results, search the literature, and carry out multi-step research tasks with limited supervision. Between those two dates, the useful human skill shifted. It is no longer prompting — asking nicely, in the right words, is a solved problem. The skills this chapter teaches are specification (stating a task precisely enough that success is checkable) and evaluation (checking it).
Nothing here contradicts Chapters 1 through 5; it depends on them. An agent’s output is an untrusted analysis, and you already know what to do with an untrusted analysis: demand a reproducible environment (Chapter 5), a fair evaluation against a held-out truth (Chapter 3), and a baseline it must beat. This chapter applies that machinery to the agents themselves.
A note on how we work in this chapter: the notebooks do not call any external AI service. They run against provided artifacts — recorded agent answers, simulated agents with planted failure modes — so they execute in CI and cost nothing. The reasoning transfers directly to the live agents you use in your project, where the course AI policy applies: use is allowed, disclosed, and verified, and you must be able to defend every line you submit (Chapter 1.8).
What is in this chapter¶
- 6.1 From language models to agents — Concepts: what a transformer-based LLM is and is not, why LLMs are bad calculators, retrieval, tool use, and hallucination as a failure mode distinct from ordinary error.
- 6.2 Critical evaluation of AI output — Hands-on: verify an agent’s quantitative claims against data, catch a fabricated citation, and score two AI reviews to see judge biases in action — then swap scores with a partner to measure whether two raters applying the same rubric even agree.
- 6.3 Build an evaluation set — The core graded exercise of the chapter: before trusting an agent with a task, build the eval set that measures it. Worked end to end on a GNSS velocity task — spec, cases, a scorer that survives malformed output, tolerances derived empirically from synthetic realizations — then extended in “Scoring without computable truth” to tasks with no exact answer: rubric-as-scorer, two-rater agreement (percent and Cohen’s kappa), judge-bias controls, and pass rates over repeated trials of a stochastic agent. An optional closing section runs the same harness against a live open-weights model (OLMo 2 via Ollama); the offline simulated/recorded path remains the graded baseline. The exercise repeats on a task from your own project domain.
- 6.4 Disclosure and norms — Attribution and disclosure for AI-assisted research: what journals expect, the course disclosure format, the institutional layer beyond the journal (employer/sponsor policy, data classification, transcript retention), and who owns correctness (you).
- 6.5 The reading arc — The quarter-long capstone contract: an AI-assisted literature review with a citation-verification log, the anatomy of good papers, a quality rubric you write in your declared genre, and a pre-submission review agent built from that rubric and evaluated with this chapter’s machinery.
Learning outcomes¶
By the end of this chapter you can:
- explain what an agent is (an LLM, a set of tools, and a loop over observations) and where in that loop errors enter;
- verify an AI-generated claim against the underlying data rather than against your impression of plausibility;
- design and run an evaluation set with ground truth for an agent task before deploying the agent;
- score an agent whose output is a judgment rather than a number: rubric checks over planted defects, rater agreement, and pass rates over repeated trials;
- disclose AI assistance in the format the course, current journal policies, and your institution require;
- turn your own quality standards into a review agent you have measured, and defend both the standards and the measurement.