Three judge-verified samples of the same builder screen, built through the OpenDesign pipeline from community research (n8n, Zapier, Make, Gumloop, OpenAI AgentKit, AgentHub). Same workflow underneath: abandoned-cart recovery with typed edges, approval gates, compiled code. Pick the one that feels most trustworthy to you.
The full three-column studio: Copilot proposes graph diffs (ghost nodes, Apply/Discard), typed-edge chips, Check panel, versions & templates. You get: per-node run badges, insert-on-wire ⊕, minimap, zoom, Check panel with type-mismatch/warning, run-to-here ▶, palette strip. Cost: three columns = densest screen; smallest merchants may need a guided mode on top.
n8n's trust model in Calm Studio clothes: every node shows its last run's in→out data, and a timeline strip shows the run as bars in the graph. Strongest for: hands-on operators who debug workflows themselves. Cost: densest learning curve; the Copilot chat is not on this screen — it's an expert console.
Gumloop-style polish: the conversation is the product; the graph is a read-only proof with pending ghost nodes and a cost prediction strip. Strongest for: non-technical merchants; lowest time-to-first-automation. Cost: direct editing is demoted to a drawer — power users lose the always-there canvas and inspector.