Designing an app · Lesson 5
An agent is the part of the app that works with judgement: it reads a contract, compares, decides what to flag. The manifest says what it is asked (the goal), what it may use (context and tools), what it gives back and where that goes, and how it works with other agents. Here the same reviewer grows from a single sentence to a small team.
Use Next, or the ← → keys. Edit the form: the task is rendered by the runtime’s own goal renderer. The tools are listed exactly as the app’s MCP server would serve them.
key · role · description — its name; main (the one the form starts), helper, scheduled, analyst, setup, designer; one line other agents read.profile — inline (crews: engine, model, system prompt, MCP servers) or ref to a profile on the platform.goalTemplate · when — the task with placeholders and optional blocks; when chooses between agents of the same role from the form.context · variables · startOptions — records or slot values put in the task; profile variables; extra fields for the start.appTools · mcpServers — the app’s tools (read, search, write, append, show, read_file, http, call); external services as MCP servers, secrets as ${VAR}.output · resultSchema · writesTo — what comes back; the shape of its structured part; the record it becomes.delegates · awaits · memory — the edges that generate ask_…, request_… and remember.limits — how often it may run, per user or plan (lesson 7).Lesson 6 is about what happens by itself: triggers, schedules, reminders, personal agents, and the human checkpoints that pause the work.