0c33/agentic-ai

Agentic AI Architect is a Python program that builds other small Python programs called agents. The user describes an idea in plain words, and the system runs through a series of loops to turn that idea into a working, standalone Python file. The README is clear that this is developer alpha, not consumer software, and requires manual setup. The README sets out a core philosophy and a system flow chart. The headline idea is that the program asks before it builds, tests before it trusts, and refines until a human approves. There are six numbered phases. Phase 1 sets up a connection to a local large language model running on localhost port 8080 using an OpenAI-compatible interface, through the LangChain library. Phase 2 is a clarification loop where the model keeps asking the user questions and storing the full chat history until it reports that the user's idea is fully understood. Phase 3 is a design loop. The model proposes an agent name, scope, prompt, and arguments, then a function called Sandbox runs the proposed prompt against the model itself to see if it gives a sensible answer. This is described as prompt validation rather than running real code. The loop keeps a history of prompts that were tried and what came back, plus a try-counter that stops it from finishing on the first attempt. A human must approve the design before the system moves on. Phase 4 generates actual Python code. It feeds the model a template, a copy of the main program's own source code as a reference, and the chosen prompt and arguments. Phase 5 then runs that generated code for real, in a separate Python subprocess from a dedicated virtual environment, with the test input piped to standard input and the output captured. A second model call rates the result on a numbered schema and decides whether to remake the code or to ask a human for final approval. Phase 6 writes the approved code to a file named after the agent. The README closes with an example of a generated class called SuperIdeaToAtomicTasks, a table of components and their purposes, and engineering notes that flag the two distinct testing stages, the self-referential code generation, the full state tracking across loops, and the use of a try-counter to prevent the design loop from exiting too early.