LangGraph vs agent loops: choosing an AI harness

Cloud Codes compares model-directed loops with workflows whose states and transitions developers define in code. The question is how much control an AI agent needs around the model, and whether that structure helps on the task at hand.

The speaker describes Mini SWE-Agent as roughly 100 lines of Python with bash as its only tool and a linear history. For comparison, a reported customer-service trial used LangGraph workflows with 14 to 55 nodes. In the travel-booking condition, the speaker reports errors in 24% of graph runs versus 11.5% for a loop using the same model and procedure. These figures describe the cited experiment rather than a general ranking of architectures.

Long tasks introduce other concerns. The discussion covers accumulated errors, declining retrieval accuracy as context grows, and checkpoints that let execution restart or resume. It also examines GPT-Researcher's reported removal of its LangGraph multi-agent pipeline and benchmark examples where changing the harness affected a coding assistant's score without changing the model.

The speaker recommends a loop when the task fits one context window and checking the answer is cheap. Explicit workflow controls become more useful when execution takes hours or must recover from interruptions. The closing advice is to reassess that structure as models improve.