Micro-Agent
A lightweight Python framework for quickly building a specialized AI agent for one industry or task and delivering it as an API service, with built-in retrieval-augmented generation and injectable domain skills.
Micro-Agent is aimed at building a professional agent for a specific vertical and delivering it as an API service in as few steps as possible, positioning itself against more general frameworks such as LangGraph and AutoGen, which the project says need more setup work to reach the same out-of-the-box API service, domain-skill injection and multiple large-language-model profile configuration.
Its core is a ReAct-style think-act-observe execution loop that supports dispatching subtasks to subagents and running code in a sandboxed REPL, a session memory system with short-term memory, file-based persistence and cross-session recovery, a skills layer that injects domain rules and coding standards into an agent's system prompt, retrieval-augmented generation over a domain knowledge base, and MCP-based tool and data-source integration. Model access goes through litellm, so one codebase can switch between OpenAI, DeepSeek, Claude or Ollama models, and different tasks can be configured with different model profiles, such as a fast profile and a reasoning profile.
A new vertical task is defined with a prompt template, registered with a task configuration naming its model profile and step limit, and then called over its REST API; it also runs in Docker.