The vision paper argues that trust in agent-to-agent networks cannot be retrofitted onto protocols designed for isolated LLM agents and must instead be architected from the initial coordination framework. Existing alignment methods fail to mitigate risks including adversarial composition, semantic misalignment, and cascading failures as heterogeneous agents scale to collaborative multi-step tasks. Four design pillars are outlined to embed trustworthiness at the network level. Related work on frameworks such as AutoGen demonstrates performance gains from agent collaboration but omits native trust mechanisms, leaving production deployments exposed to the same coordination failures identified in the arXiv submission. Studies on LLM agent ecosystems further highlight emergent misalignment patterns that single-agent techniques cannot contain once agents interact autonomously. The coverage understates the transition timeline to production systems, where these gaps compound without foundational redesign. Primary evidence from the 2026 arXiv paper is synthesized with patterns in multi-agent orchestration literature to show that bolted-on solutions consistently lag behind network-scale threats.