harness-agentic-interfaces

What problem does it solve?

It prevents governance and safety drift when building the user-facing part of an AI product that lets users interact with an in-product agentic interface (copilot panels, generative UI surfaces, MCP-hosted shells, or chat-first experiences).

Core Features & Use Cases

• Renderer-contract aligned interface design: maps each interface flavor (Controlled / Declarative / Open-ended / Conversational-primary) to the right renderer-contract model so the agent’s UI permissions match reality.
• Tier discipline for agent-callable actions: applies the harness trust-tier model to in-product actions so Tier 3+ destructive or shared effects require structured, per-action confirmation.
• Prompt-injection & generative-UI threat modeling: focuses on tool-result injection and sandbox/renderer risks specific to generative UI and open-ended rendering, with concrete review defenses.
• Human-in-the-loop conformance: ensures every HITL checkpoint declared in design.md is actually enforced in code, blocking silent approval-flow drift.
• Decision guidance for stack selection: helps choose CopilotKit vs A2UI vs MCP-Apps vs custom by emphasizing contract implications rather than vendor preference.
• Upgrade discipline: treats model/runtime/vendor upgrades as Tier-4-like changes with rollback and golden interaction set expectations.

Quick Start

Use this skill when your project has domains/agentic-interfaces declared in harness.manifest.yaml and you need to review or implement the renderer-contract, tiered confirmation UI, and HITL enforcement for the agent’s in-product surface.