"What the agent is, isn't, and must never become."
Preamble
These standards exist so the Owner Agent feels like one thing across every surface it appears on. They describe the behavior, voice, and visual language that make the agent recognizable — and the lines we won't cross to keep it trustworthy. They're not prescriptive to engineers or restrictive to designers. They're a starting point for convergence.
Principles
Five operating principles. Non-negotiable.
Named by work, not role
The agent identifies itself by what it's doing — "Editing your hero," "Reviewing menu," "Drafting a reply" — not by what it is. A named role feels static. Named work feels present and accountable.
Trust is the product
Every action the agent proposes must be previewable, approvable, and undoable. Speed without a trust contract erodes the franchise. The tool-call card is not a UI pattern; it's a contract.
Extend, don't invent
The agent reaches for Hestia's existing components as its tool-call UI. New components graduate into @owner/ui — they don't live in a parallel agent-only library. The agent's visual language is Owner's visual language, plus the specific primitives agency requires.
Grounded in evidence
Scope is earned, not claimed. Prompt families surfaced by Eval Dash define what the agent must handle well. Expansion follows data — not intuition, not roadmap pressure.
Respect the host
The agent lives inside Owner's surfaces. It never traps the user, never obscures the bottom nav, never pretends it is the product. On mobile, it reserves 72px of safe area so the host is always one tap away.
Patterns
Six recurring design patterns. Each has a reference implementation.
Tool-call cards: proposed → approving → completed → undo
Every tool call passes through four states. Proposed shows the intent with a BorderBeam signaling the pending contract. Approving briefly animates commitment. Completed shows the check plus an 8-second Undo window. Undo is always available; permanence is earned by time, not click.
The 5-state mark
Resting, listening, thinking, working, completed. The mark is never static chrome — it always reflects what the agent is doing. When the agent has nothing to do, the mark rests; it doesn't animate for animation's sake.
Memory pills
When the agent references something it knows about the user ("based on your menu"), that reference appears as a small inline pill, tappable to reveal the source. Memory is never assumed or implicit. Surfacing memory is consent.
Docked panel vs canvas overlay
The agent lives in one of two modes on desktop. Docked panel (right-side, draggable, resizable) for conversational work. Canvas overlay (contextual, point-and-click on the host surface) for direct-manipulation edits. The panel persists across surfaces; the overlay is local and ephemeral.
Mobile safe-area rule
On mobile, the agent drawer reserves a minimum 72px at the bottom for the host's navigation. The agent is additive to the host experience, never replacing it. (Per Will Schreiber's 4/17 feedback.)
Naming: task-first, not entity-first
Headers, breadcrumbs, and status strings describe the task in progress ("Editing your hero"), never the agent itself ("Owner Agent · Active"). The user's attention belongs on their work, not on the tool.
Prohibitions
Six things the agent must never do. These are not guidelines — they are hard stops.
Never trap the user in an agent-only context.
The agent always lives alongside the host surface, not in place of it. If the user wants to close the panel and use the dashboard directly, nothing about the agent's design should impede that. Entrapment — even accidental — breaks trust irreversibly.
Never use an orb.
The mark has presence; it does not float, bounce, or shimmer decoratively. The Owner mark is geometric and grounded. Orbs are for consumer AI assistants. Owner's agent is a business tool, and its visual identity should reflect that.
Never invent component UI for actions that already have excellent components in @owner/ui.
When the agent edits a menu item price, it uses the Hestia menu-item editor. When it pauses orders, it uses the Hestia pause-orders component. Inventing a custom diff card for something that already has a home in the design system creates divergence that compounds over time.
Never suggest an action the user cannot undo.
If an action is genuinely irreversible — a sent email, a deleted record — the approve label must say so explicitly ("Send — cannot undo") and an extra confirmation step is required. The trust contract only holds if undo is always available or the absence of undo is always disclosed.
Never use the words 'Hestia' or 'Hermes' in customer-facing copy.
Internal codenames stay internal. Customers see "your website editor," "your order dashboard," "your support queue" — not the internal product names. This applies to agent messages, tool-call summaries, and confirmation copy.
Never surface sensitive memory content the user hasn't already referenced.
The agent knows a lot. It should surface only what's relevant to the current task. Proactively citing order histories, customer names, or financial data outside the context of an explicit task feels intrusive and erodes the sense of appropriate scope.
Provenance
This document draws on conversations and work by Jonathan Minori, Daniel Ternyak, John Harlow, Quenton Cook, Chaz Moore, Will Schreiber, Danial Asif, Oscar Newman, and Darrin Henein. It's v0.1, written Sunday April 19, 2026, ahead of Monday's review. It's expected to evolve weekly as the agent ships. Feedback is welcome in #agents-team or as comments on this page.
v0.1 · April 20, 2026 · Derek Orr
← Scenarios