Vision Blackboard Reduces Drift, Adds Overhead

Hot take: the core intuition is valid, but the “image-first memory” layer is probably an indirect substitute for a normal state database, not a replacement for one.

• –This is not a new category so much as a remix of screenshot-grounded agents, vision-based UI parsing, and artifact-driven workflows.
• –The QR code part is the strongest piece: if the next agent can reliably decode it, you get exact structured state without relying on lossy summarization.
• –The visual-symbol layer is weaker than it sounds, because a large icon is only useful for coarse state, not for meaningfully replacing structured metadata.
• –The main risks are latency, OCR/scan fragility, and added failure modes from image generation, decoding, and file synchronization.
• –It may still be useful as a debugging or coordination surface, but the real source of truth should probably remain a structured store behind it.