Books / Learning comic
AI Agents: From Instructions to Useful Work
Theme: Computing & AI
Mira, a facilities manager, and Jonas, an engineer, turn Scout’s polished but unsupported recommendation into a checkable research draft. Across eight illustrated chapters, they define a six-field comparison, inspect versioned evidence, recover a timed-out save, test exact-content approval and let a simulated sharing failure change the release decision. The story ends with a reviewed local brief and a bounded draft-only pilot. A 29-page supplement contains all twelve fictional records, exact document versions, recovery and approval histories, evaluation outcomes and worked checks. Page-linked notes explain the mechanisms and provide questions with answers. All people, data and outcomes are invented for teaching.
Start reading →View in the interactive libraryAbout this book
Mira, a facilities manager, needs an evidence-linked comparison of three community buildings. Jonas, an engineer, helps her build Scout, a fictional research assistant. Its first attractive recommendation has no supporting source. They replace it with a bounded task: compare documentation readiness under six explicit requirements and produce a local draft for human review. Eight illustrated chapters follow the consequences of that choice. The team counts usable fields, preserves missing values and checks each claim against a versioned source. A tempting two-day disruption estimate turns out to describe a rejected option. A timed-out save and interrupted receipt update require reconciliation of the same operation, without a duplicate draft or a reset budget. The evidence, unknowns and remaining resources carry through one continuous research job. An isolated sharing experiment binds approval to an exact document, recipient and permission. A changed attachment exposes a gap between the review screen and the execution path. Seventeen passing tasks out of twenty cannot outweigh that simulated unauthorized effect. Sharing remains disabled while the team diagnoses the failure, tests a focused repair and separately establishes the gates for a draft-only pilot. The final evidence-linked brief and a new broken-plan capstone bring the lessons together. The eight chapters total 83 pages. The 29-page supplement, Inspectable Records and Worked Checks, supplies the complete twelve-record source packet, exact v1/v2 documents, approval and recovery histories, all twenty evaluation outcomes, worked calculations and linked primary references. Page-linked notes add explanations, specific hints and answered questions. All people, records, rates and outcomes are fictional teaching fixtures. The example establishes no deployed-system reliability, certification, engineering suitability, energy savings or authority for external action.