Agent Loop Builder

The canvas begins with a system prompt, model, tool and evaluator: four nodes and four edges. Removing a node deletes every incident edge, and dragging changes only visual positions. The output is a Python code scaffold. The browser never launches an agent, executes Python or calls a model when you edit this graph. The variable called has loop tests only whether any edge exists. It does not detect a directed cycle or compile the exact topology. One connection is enough to emit the same fixed loop, and no connections emit a pass placeholder. The first system node controls a static prompt and its cache flag adds ephemeral cache control. The evaluator node never compiles an evaluator implementation. The scaffold sets a maximum of ten iterations and a per-request token parameter of one thousand twenty four. Its model identifier is a fixed historical string, not a verified current endpoint. Tool execution returns the literal success message and evaluator logic is a comment. The generated provider call is text only and was never run here. Cache control does not prove actual cache hits or cost savings. This local preview preserves the original lifecycle allocation gate, with external requests aborted for checks.