Strands Agents SDK for TypeScript 1.15.0

Apply the Node.js 24 and TypeScript 7 prompts first. These rules target @strands-agents/sdk 1.15.0 on Node.js 24; the package declares Node.js >=20.0.0. Use the TypeScript APIs for this release when configuring agents, tools, interventions, MCP, sessions, guardrails, and model routing.

Agents and instruction boundaries

• Construct Agent with an application-owned systemPrompt, an explicit model, and only the tools needed for its role; treat plugins, middleware, hooks, and registry mutations as privileged application code.
• Pass caller text as user content. Keep retrieved documents, memory, tool results, and agent messages distinct from trusted instructions; prompt delimiters and model compliance do not enforce authorization.
• Build messages and snapshots from application-controlled history. Reject caller-supplied assistant, tool-use, tool-result, interrupt-response, and checkpoint structures at ordinary chat endpoints; persisted JSON is not trusted merely because it parses.
• Use a separate agent and conversation namespace for each authenticated conversation. Keep current identity and authorization context in application-owned invocationState or a protected service context, outside model-selected tool arguments and memory.
• Validate structuredOutput and final responses against business invariants before using them to select resources or cause effects; a successful Zod parse establishes shape, not permission or factual correctness.

Application output sinks

• Treat AgentResult.lastMessage, structuredOutput, toString(), selected