Engineer

The Engineer is the code-writing agent in NodeGPT's stable of AutoGen personas, and it's the star of the pack's flagship demo: a workflow called Task_Solving_with_Code_Generation that takes a task, has the Engineer write Python to solve it, and then actually runs it. If you've seen the screenshots with a chat that appears to write and debug code by itself, this is the node behind the "writing" half.

Like every agent node in the pack, it's a thin wrapper around Microsoft's AutoGen with a persona baked in. You feed it an LLM and get back an Agent that other nodes can chat with. The differentiation from the pack's other agents - Assistant, Critic, Planner, Executor - is all in the system message: Engineer is told to produce code, Critic is told to review it, Executor is told to run it. That division of labor is the entire architecture, and it's a genuinely good design pattern that survived this pack's demise.

Inputs and output

• LLM - required, from any provider node (ChatGPT, LM_Studio, Ollama, llama-cpp, oobaboogaOpenAI).
• Seed / Temp - default 0. Keep Temp low for code; you want reproducible output, not creative license.
• request_timeout - 120s default; bump it when the model is slow and writes long functions.

Output: Agent, wired into a GroupChat with a Critic and Executor, or into a Chat node.

The part people actually hit

Community reports from the era describe the Engineer (and the agents after it) "just repeat[ing] whatever is passed to them" - a failure mode that reads as the persona not sticking, usually downstream of the temp/seed setup or a local-model quirk on the server side. And the pack's own docs are blunt: "not everything is functioning correctly yet." The author also scoped the demo explicitly to code generation with a human in the loop; this was never "AI writes and runs anything," which is probably wise given the Executor node exists to do exactly that.

The repo is now deleted - xXAdonesXx/NodeGPT 404s as of this writing, author's account gone - so the README's install is historical:

cd ComfyUI/custom_nodes
git clone https://github.com/xXAdonesXx/NodeGPT

Restart to auto-install requirements (including pyautogen). If you're adapting the idea today, the takeaway that aged well: split "write" and "run" into separate agents with separate system messages. That separation is what keeps a coding agent from being a footgun.