harness-agentic-interfaces

Safe in-product agentic interfaces require mapping each UI flavor—Controlled, Declarative, Open-ended, or Conversational-primary—to a matching renderer-contract model so agent permissions align with reality. You must apply tier-aligned structured confirmations for Tier 3+ actions and conduct prompt-injection threat modeling.

Human-in-the-loop checkpoint enforcement requires declaring every HITL checkpoint in design.md and blocking silent approval-flow drift in code. You must apply the harness trust-tier model so Tier 3+ destructive or shared effects trigger structured, per-action confirmation before execution.

Prompt-injection threat modeling for generative UI focuses on tool-result injection and sandbox or renderer risks specific to open-ended rendering. It provides concrete review defenses to prevent unauthorized agent actions and ensures renderer-contract permissions match the actual UI capabilities.

Choosing between CopilotKit, MCP-hosted shells, A2UI, or custom stacks should emphasize contract implications rather than vendor preference. The decision depends on which framework best aligns your interface flavor—Controlled, Declarative, Open-ended, or Conversational-primary—with the correct renderer-contract model.

Structured, per-action confirmation is required for all Tier 3+ agent-callable actions that produce destructive or shared effects. This tier discipline applies across copilot panels, generative UI surfaces, MCP-hosted shells, and chat-first experiences to prevent unsafe autonomous execution.

Keeping design.md and code in sync requires HITL checkpoint conformance verification that blocks silent approval-flow drift. You must ensure every declared human-in-the-loop checkpoint is actively enforced in the renderer-contract and tiered confirmation UI implementation.