AgentSwarms adds image playground, vision loops

This is less about image generation itself and more about making multimodal agent orchestration usable without writing glue code. The feature targets a real pain point: once you add critique loops and vision feedback, the routing and debugging complexity explodes.

• –The visual canvas abstracts the worst part of multimodal systems: handoff logic between prompt, generation, and critique steps
• –Vision-in-the-loop workflows are useful for self-correcting outputs, but they can get expensive fast if the retry logic is sloppy
• –Real-time payload inspection matters here because multimodal failures are often hidden in intermediate transforms, not final outputs
• –This positions AgentSwarms as a learning and prototyping layer for agentic creative workflows, not just a generic chat sandbox
• –The main question is whether the workflow designer stays simple enough once teams start adding more nodes, branches, and evaluation rules