Návrh a testování běhových prostředí pro autonomní coding agenty v softwarovém vývoji

AI coding agents can autonomously generate code, tests, and other development artifacts, but their practical usefulness depends on more than whether the resulting program passes tests. How the agent worked, whether it left an auditable trace, and the quality of the resulting code also matter. Existing benchmarks rarely capture these dimensions together, and no established process exists for designing instructions against such measurements. This thesis proposes a metric suite covering process, product quality, and resources, and shows how it can guide iterative instruction design. Its feasibility is evaluated on a case study of a billing reminder system: the agent repeatedly implements the same specification, each run is evaluated by the proposed metrics, and the instructions are revised based on diagnosed failures. Subsequent ablations examine which instruction components contribute measurably. Results show that the metrics can distinguish a functionally successful output from one produced by a weak or non-auditable process. Iterative instruction changes improved agent behavior, though not steadily: some changes caused regressions, and identical instructions produced different process adherence across runs. Successful changes followed a recurring pattern: a shift from general rule to specific command to verification step. Ablations showed that verification steps are not redundant; removing part of the code-convention section had little effect on deterministic metrics but reduced design quality assessed by an LLM-as-judge. The contribution is a feasibility evaluation of the metric suite and instruction-design process on a single case. Specific measured values are bound to the model, tool, and project; what transfers is the process and the metric suite. Whether the observed operationalization pattern holds outside programming remains an open question.