Agentic AI: The Agent Engineer's Stack

The author, a Habr user, notices that many candidates for agentic Python engineer positions struggle with common concepts and tools in the field. To help, he provides a practical overview. He defines an agent as an entity with a prompt, an LLM, and tools, but crucially with a loop and feedback: the model decides each next step, acting and observing results, unlike a pre-defined chain. He notes that while chains have a fixed route, agents have a set of capabilities and a goal, with the route emerging dynamically. The loop involves tool declarations via API schema (JSON Schema), the model requesting tool calls, and the runtime executing them and returning results to the context. Stop conditions, such as the model deciding the task is complete or iteration limits, are important. PydanticAI is highlighted as a type-safe framework for building agents, using Pydantic models for output validation, tools with type annotations and docstrings, dependency injection via RunContext, and output validators. LangGraph is presented as a graph-based approach for stateful workflows, allowing human approval and checkpoints. smolagents from Hugging Face uses code agents that write Python code for execution. CrewAI offers a role-based approach with agents having roles, goals, and biographies, orchestrated in a crew. The article also mentions MCP (Model Context Protocol), an open standard by Anthropic for connecting tools, described as USB-C for AI, enabling interoperability. The material is aimed at helping readers pass interviews, acknowledging that the field evolves rapidly.