AI Architecture Updates: July 14, 2026

1. The Harness Around the Model Is the Real Enterprise Agent Architecture

ByteByteGo. The newsletter argues that production agents at scale depend less on the model and more on a surrounding harness of inference, runtime, observability, identity, and context layers. Two patterns stand out: treating retrieval as an agentic subagent that plans queries, evaluates results, and recovers from failed lookups rather than doing one-shot RAG, and giving action-taking agents their own organizational identities with audit trails bounded by the same access controls as human employees. It also pushes evaluation from pre-deployment testing toward continuous, rubric-based assessment of specific expected behaviors instead of generic quality metrics. Source

2. DoorDash Keeps Business Logic Out of the Prompt With a Runtime and MCP Layer

InfoQ. A case study on DoorDash’s Ask Assistant describes an Assistant Runtime that orchestrates specialized agents rather than embedding logic into prompts, backed by a Model Context Protocol layer that exposes reusable tools over existing services like catalog search and recommendations. Memory is split into three tiers, long-term behavioral history, session conversational context, and explicit agentic facts, retrieved by semantic search and ranked before entering the prompt. The design leans on deterministic actions that update versioned artifacts without invoking the language model, which the team credits for faster and more reliable behavior than an LLM-only approach. Source