A single prompt-and-response is fine for a question. It's the wrong shape for "brainstorm this, then turn it into code, then clean up the explanation." Playground is Trunk's answer: chain reusable Agents into a linear Flow, and run the whole thing on-device, no cloud round-trip between steps.
An Agent is deliberately small: a name and a system prompt, nothing else. Wire a few together and each one only has to be good at its own step. Below, a "Brainstorm" agent's output becomes a "Code writer" agent's input, whose output becomes a "Formatter" agent's input, three narrow jobs standing in for one that would otherwise be too vague to prompt well. Tap any node to read the exact prompt driving that step.
Name the agent, write its system prompt directly. Save it once, reuse it across as many Flows as you want.
Describe the agent's purpose in plain language and the currently-loaded model writes the system prompt for you, entirely on-device. Review and edit before saving; nothing is ever auto-committed.
Every agent in a Flow shares whichever model is currently loaded, an agent is a different system prompt, not a different model download, so switching between them is instant.
Drop Agents onto a canvas, drag them into place, then tap an output handle followed by an input handle to connect two nodes in order. No YAML, no node-graph SDK, just touch. Assign a model and generation settings to the Flow once, then run it whenever the input changes.