PROFESSIONAL PRACTICE AND COLLABORATION

PLANNING AND ESTIMATING HONESTLY Work on agents is harder to estimate than ordinary software, because results depend on how a model behaves on real data. A prototype may take a day, while reaching dependable quality may take months. When you estimate, separate three things: building the feature, measuring its quality, and hardening it against failures. Give ranges, not single numbers, and say what could make the range larger. Show early results often, because a rough demo with honest scores helps everyone steer sooner. When a deadline is firm, reduce the scope and not the testing, since untested agents cause the most expensive surprises. CHOOSING WHAT TO BUILD AND WHAT TO RENT Every team must decide which parts of a harness to build and which to buy or adopt. Chapters 49 through 52 described frameworks, standards, and managed platforms. A sensible rule is to build the parts that make your product special and to adopt the parts that are common to everyone. Weigh the cost, the risk of being locked in to one vendor, the control you need over safety, and the effort to maintain the choice. Write the decision and the reasons in a design note so that the choice can be revisited when circumstances change.