Vercel AI SDK makes tool loops a frontend-native agent primitive

Agent products increasingly need the same primitives inside web apps, API routes, terminal tools, and coding-agent harnesses. Vercel's AI SDK Agents documentation defines agents as LLMs that use tools in a loop, then exposes that pattern through `ToolLoopAgent`: typed tools, shared `runtimeContext`, per-tool `toolsContext`, configurable `stopWhen` loop control, `toolApproval` for sensitive operations, terminal UI support, and `HarnessAgent` for established coding-agent harnesses such as Claude Code, Codex, or Pi. The companion tool-calling documentation also describes strict schemas, dynamic descriptions, tool results, and approval request/response flows. For agent builders, the practical signal is that frontend-native agent stacks should be judged by whether the loop is inspectable, typed, state-aware, approval-aware, reusable across UI/API surfaces, and still able to fall back to explicit structured workflows when determinism matters.