The emergence of AutoML technology has lowered the technical threshold for developing AI, but current systems address merely separate phases in the ML lifecycle and require advanced configuration skills and lack transparency in their architectural decisions. Most frameworks assume running the pipeline on the cloud or containerized environment, which restricts their applicability in privacy-preserving and resource-poor scenarios. In this work, we introduce the AgentDLOps, an autonomous agentic system that incorporates the following elements into a cohesive end-to-end ML lifecycle management solution: LangGraph state machine with 13 nodes, large language model (LLM) reasoning component (Llama-3.3-70B), Jinja2-templating local code generation with AST-level static safety checks, and real-time React dashboard. With just a natural language problem description and CSV dataset provided, AgentDLOps performs data profiling, architecture selection, generates and validates the training script, runs the training script locally, evaluates the result in regard to the set accuracy target, optimizes iteratively with the predefined improvement budget, and finally outputs a model artifact and metrics per class in a structured manner. The core principle is: LLMs reason, tools execute – this is achieved via template and AST-level safety checks. On the retail product classification benchmark (85,000 samples, 27 classes), we get a test accuracy of 83.83% with 13.56 training time and 32.3 s total pipeline time. On the Iris benchmark, we reach an accuracy of 96.67% in 0.41 s. AgentDLOps is within 0.27% of the performance of a hand-crafted expert baseline while removing all need for user-side engineering work and beating plain LLM code generation in pipeline success rate by 29%.