// the take
PraisonAI is a Python framework for building autonomous LLM agents and multi-agent workflows, bundling prompting, context/memory management, tool and MCP dispatch, iteration control, and workflow graphs (route/parallel/loop/repeat) behind one API, plus optional remote sandboxes for running tools or whole agents. It's for Python developers who want more structure than a hand-rolled agent loop but don't want to assemble LangChain/LangGraph-style pieces themselves.
The five-layer breakdown (prompt/context/harness/loop/graph) maps directly onto concrete parameters (memory=, tools=, execution=, AgentFlow), so when an agent misbehaves you have a specific layer to inspect instead of one opaque orchestrator. It ships real guardrails against runaway loops out of the box: max_iter, max_budget with on_budget_exceeded, and doom-loop detection for repeated or A/B-oscillating tool calls, things most people only add after burning API spend the hard way. MCP support spans stdio, HTTP, WebSocket and SSE, and tools_run_on=/run_on= let you move just the tools or the entire agent to Docker, e2b, Modal, etc. without rewriting agent code. YAML and Python both compile to the same underlying agent graph, so trivial pipelines don't need code but you can drop into Python when YAML runs out of room.
100+ LLM providers plus dozens of RAG backends, vector stores, messaging channels and external CLI integrations (Claude Code, Gemini CLI, Codex) is a lot for one project to keep correct — expect uneven provider coverage and wrappers that lag behind the SDKs they wrap. The memory/knowledge/RAG stack defaults to 'zero deps' implementations that are fine for a demo but will need swapping for a real vector store in production, and claims like quality-scored retrieval or graph memory need independent verification before you rely on them. The README itself warns against eval()/exec()/subprocess in tool functions, which signals that unsafe code execution is an easy trap with this tool model — treat sandboxing as mandatory, not a nice-to-have, for anything touching user input. Feature velocity is very high (managed sandboxes, A2A protocol, bot gateway, doom-loop detection all layered on recently) from what looks like a small core team, so check issue/PR responsiveness before betting production agent infrastructure on APIs that may still be moving.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →