Chapter C75 · Interactive Course
Useful agent work beyond a demo
Research briefs, data cleanup, rendered documents, live interfaces, automation limits and the simplest adequate system.
Core Learning Concept
Write a personal use-case brief for one achievable task. State the desired artifact, permitted inputs and actions, required checks, stopping rule, and recovery route. Choose the simplest suitable workflow and try a small version. The course ends with a process you can inspect and improve in your own work.
What You’ll Learn in This Chapter
-
A research brief needs traceable evidence
Define a report task with source requirements.
A research brief connects factual claims to primary sources readers can inspect. State the scope and dates clearly, separating evidence from interpretation. Fluent summaries lacking verifiable citations leave readers unable to assess truth or reliability.
-
Data cleanup needs preserved meaning
Specify a transformation using a tiny dataset.
Data cleanup should preserve meaning while changing representation. Trimming whitespace is different from merging people with similar names or guessing missing values. Define allowed transformations, keep originals recoverable, and compare totals or representative records. Flag ambiguous cases instead of silently inventing a tidy-looking answer.
-
Document production needs output inspection
Define a usable artifact beyond generated text.
Document production includes creating the file and inspecting its rendered result. Check headings, tables, page breaks, figures, accessibility, and content accuracy for the intended reader. A file that opens successfully can still be unusable. Review the exported artifact rather than relying only on the source instructions.
-
Browser tasks need state checks
Use observations before actions on a changing interface.
Browser tasks operate on changing interface state. Before an action, confirm the page, selected account or item, and control being used; afterward, inspect the result. Loading, navigation, and duplicate submissions can change the situation. A click is evidence of an attempted interaction, not necessarily a completed task.
-
Routine automation needs a recovery plan
Choose a stable repeated task with clear failure handling.
A recurring automation needs a defined trigger, scope, success signal, and recovery behavior. Consider partial completion and repeated side effects before enabling it. Decide when to notify a person and how access can be stopped. A workflow that works once still needs handling for later failures and changed inputs.
-
Choose the simplest adequate system
Compare a fixed script, an assistant and a persistent agent.
Choose the simplest system that meets your goal with acceptable effort. Simple shell scripts, standard tools, or assisted workflows often outperform complex agents. Factor in setup, review, and recovery costs rather than favoring impressive demonstrations.
-
Make a personal use-case brief
Specify inputs, permitted actions, output evidence and a stopping condition for a research, data, document or coding task.
-
Choose an agent for a real need
Compare candidate tools on the same small task and justify a choice using evidence, access and workflow fit.