From LLM Prompts to Controlled Execution: The Experience of Developing an AI Agent / Habr
TAG62I would note: the environment has been deliberately simplified to constrain the experimental conditions and analyze the model's behavior more accurately.

1.2 How the project evolved. Brief chronology

TAG69TAG72TAG79

TAG81@@Stage

TAG84TAG103

Design of the synthetic environment, states, contracts, tools, checks, trace, and evidence.Deterministic testing of the first scenario, S01, without an LLM.

TAG74No.TAG86Brief descriptionTAG90TAG92TAG941TAG98TAG100@@Basic runtime logic

2

First LLM tests

Five single-step scenarios: diagnostics, retrieving a fact, clarifying parameters, and preparing a rollback. Comparing model quality and runtime control.

3

Deep dive into scenario S12

Investigating proposal quality, identifying an evaluator issue, changing the evaluation process, and repeating the experiment.

4

Multi-step execution

One action per step, new observations, confirmations, result verification, and safe stopping.