diff --git a/README.md b/README.md
index 6d1faddcf..6990f5824 100644
--- a/README.md
+++ b/README.md
@@ -103,6 +103,8 @@ uv add langwatch-scenario pytest
 pnpm install @langwatch/scenario vitest
 ```
 
+The TypeScript package targets [AI SDK](https://ai-sdk.dev) 7. If your project is still on AI SDK 6, stay on `@langwatch/scenario` 1.x until you upgrade.
+
 Now create your first scenario, copy the full working example below.
 
 <details>
@@ -198,13 +200,8 @@ describe("Vegetarian Recipe Agent", () => {
     call: async (input) => {
       const response = await generateText({
         model: openai("gpt-4.1-mini"),
-        messages: [
-          {
-            role: "system",
-            content: `You are a vegetarian recipe agent.\nGiven the user request, ask AT MOST ONE follow-up question, then provide a complete recipe. Keep your responses concise and focused.`,
-          },
-          ...input.messages,
-        ],
+        instructions: `You are a vegetarian recipe agent.\nGiven the user request, ask AT MOST ONE follow-up question, then provide a complete recipe. Keep your responses concise and focused.`,
+        messages: input.messages,
       });
       return response.text;
     },
diff --git a/docs/docs/pages/advanced/remote-traces.mdx b/docs/docs/pages/advanced/remote-traces.mdx
index 37a7c8c3a..a69cb4411 100644
--- a/docs/docs/pages/advanced/remote-traces.mdx
+++ b/docs/docs/pages/advanced/remote-traces.mdx
@@ -238,13 +238,23 @@ async def test_remote_trace_propagation():
 ```typescript [typescript]
 import { createServer, Server } from "http";
 import { openai } from "@ai-sdk/openai";
+import { OpenTelemetry } from "@ai-sdk/otel";
 import scenario, { AgentRole, type AgentAdapter } from "@langwatch/scenario";
 import { context, propagation } from "@opentelemetry/api";
 import { W3CTraceContextPropagator } from "@opentelemetry/core";
-import { generateText, ModelMessage, stepCountIs, tool } from "ai";
+import {
+  generateText,
+  ModelMessage,
+  isStepCount,
+  registerTelemetry,
+  tool,
+} from "ai";
 import { describe, it, expect, beforeAll, afterAll } from "vitest";
 import { z } from "zod/v4";
 
+// AI SDK 7 emits no spans until a telemetry integration is registered.
+registerTelemetry(new OpenTelemetry());
+
 // Standard OTel HTTP instrumentation registers a propagator on its own;
 // this hand-rolled demo server registers the W3C propagator explicitly.
 propagation.setGlobalPropagator(new W3CTraceContextPropagator());
@@ -277,17 +287,12 @@ beforeAll(async () => {
       const result = await context.with(extracted, () =>
         generateText({
           model: openai("gpt-5-mini"),
-          messages: [
-            {
-              role: "system",
-              content:
-                "You are an order support assistant. Use the lookup_order_status tool to check orders before answering.",
-            },
-            ...messages,
-          ],
+          instructions:
+            "You are an order support assistant. Use the lookup_order_status tool to check orders before answering.",
+          messages,
           tools: { lookup_order_status: lookupOrderStatus },
-          stopWhen: stepCountIs(3),
-          experimental_telemetry: { isEnabled: true },
+          stopWhen: isStepCount(3),
+          telemetry: { isEnabled: true },
         })
       );
 
diff --git a/docs/docs/pages/agent-integration.mdx b/docs/docs/pages/agent-integration.mdx
index 462afc675..531ad285e 100644
--- a/docs/docs/pages/agent-integration.mdx
+++ b/docs/docs/pages/agent-integration.mdx
@@ -379,7 +379,7 @@ import { z } from "zod";
 
 const getCurrentWeather = tool({
   description: "Get the current weather in a given city.",
-  parameters: z.object({
+  inputSchema: z.object({
     city: z.string().describe("The city to get the weather for."),
   }),
   execute: async ({ city }) => `The weather in ${city} is sunny!`,
diff --git a/docs/docs/pages/agent-integration/vercel-ai.mdx b/docs/docs/pages/agent-integration/vercel-ai.mdx
index df5f38dad..1b8ea1631 100644
--- a/docs/docs/pages/agent-integration/vercel-ai.mdx
+++ b/docs/docs/pages/agent-integration/vercel-ai.mdx
@@ -20,11 +20,8 @@ import { generateText } from 'ai';
 async function myAgent(prompt: string) {
   const result = await generateText({
     model: openai('gpt-4-turbo'),
+    instructions: 'You are a helpful AI assistant.',
     messages: [
-      {
-        role: 'system',
-        content: 'You are a helpful AI assistant.',
-      },
       {
         role: 'user',
         content: prompt,
@@ -46,7 +43,7 @@ const createMyAgentAdapter = (): AgentAdapter => {
     name: "VercelAIAgent",
     call: async (input) => {
       const result = await myAgent(input.messages.at(-1)?.content);
-      return result.response.messages;
+      return result.responseMessages;
     },
   };
 };
diff --git a/docs/docs/pages/basics/scripted-simulations.mdx b/docs/docs/pages/basics/scripted-simulations.mdx
index c378d8214..a4f218997 100644
--- a/docs/docs/pages/basics/scripted-simulations.mdx
+++ b/docs/docs/pages/basics/scripted-simulations.mdx
@@ -257,7 +257,8 @@ const script = [
     ],
   }),
 
-  // System message
+  // System message (with AI SDK 7, an agent that forwards these messages to
+  // generateText needs `allowSystemInMessages: true`)
   message({
     role: "system",
     content: "The user is now in a hurry and needs quick responses",
diff --git a/docs/docs/pages/examples/multimodal/multimodal-images.mdx b/docs/docs/pages/examples/multimodal/multimodal-images.mdx
index 3e8553482..a6f629bfc 100644
--- a/docs/docs/pages/examples/multimodal/multimodal-images.mdx
+++ b/docs/docs/pages/examples/multimodal/multimodal-images.mdx
@@ -36,13 +36,8 @@ const imageAgent: AgentAdapter = {
   async call(input) {
     const response = await generateText({
       model: openai("gpt-4o"),
-      messages: [
-        {
-          role: "system",
-          content: `You are a helpful assistant that can describe images.`,
-        },
-        ...input.messages,
-      ],
+      instructions: `You are a helpful assistant that can describe images.`,
+      messages: input.messages,
     });
     return response.text;
   },
diff --git a/docs/docs/pages/introduction/getting-started.mdx b/docs/docs/pages/introduction/getting-started.mdx
index 279fb0b95..7e79d85b8 100644
--- a/docs/docs/pages/introduction/getting-started.mdx
+++ b/docs/docs/pages/introduction/getting-started.mdx
@@ -26,6 +26,40 @@ pnpm install @langwatch/scenario vitest
 
 :::
 
+<Callout type="info">
+The TypeScript SDK targets [AI SDK](https://ai-sdk.dev) 7 and Node.js 22.12 or later. If your project is still on AI SDK 6, stay on `@langwatch/scenario` 1.x until you upgrade.
+</Callout>
+
+### AI SDK telemetry (TypeScript)
+
+AI SDK 7 emits no spans until a telemetry integration is registered, so the judge and LangWatch cannot see your agent's LLM calls by default. Register the OpenTelemetry integration once at start-up, for example in a Vitest setup file:
+
+```bash
+pnpm install @ai-sdk/otel
+```
+
+```typescript
+// vitest.setup.ts
+import { OpenTelemetry } from "@ai-sdk/otel";
+import { registerTelemetry } from "ai";
+
+registerTelemetry(new OpenTelemetry());
+```
+
+To attach per-call metadata to the spans (a user id, a tenant, a feature flag), pass it as `runtimeContext` and opt each key in with `telemetry.includeRuntimeContext`. Turn on `runtimeContext` on the integration too; the values are emitted as `ai.settings.context.<key>`:
+
+```typescript
+registerTelemetry(new OpenTelemetry({ runtimeContext: true }));
+
+await generateText({
+  model: openai("gpt-5-mini"),
+  instructions: "You are a vegetarian recipe agent.",
+  messages: input.messages,
+  runtimeContext: { userId: "user-123" },
+  telemetry: { functionId: "recipe-agent", includeRuntimeContext: { userId: true } },
+});
+```
+
 ## Quick Start
 
 Create your first Scenario test. This example demonstrates testing a vegetarian recipe agent:
@@ -138,10 +172,8 @@ const vegetarianRecipeAgent: AgentAdapter = {
   call: async (input) => {
     const response = await generateText({
       model: openai("gpt-4.1-mini"),
-      messages: [
-        { role: "system", content: "You are a vegetarian recipe agent." },
-        ...input.messages,
-      ],
+      instructions: "You are a vegetarian recipe agent.",
+      messages: input.messages,
     });
     return response.text;
   },
diff --git a/docs/docs/pages/testing-guides/mocks.mdx b/docs/docs/pages/testing-guides/mocks.mdx
index 3a5e94991..f104d4d69 100644
--- a/docs/docs/pages/testing-guides/mocks.mdx
+++ b/docs/docs/pages/testing-guides/mocks.mdx
@@ -161,7 +161,7 @@ const mockWeatherTool = vi.fn().mockResolvedValue("Sunny, 75°F");
 
 const weatherTool = tool({
   description: "Get weather data",
-  parameters: z.object({ city: z.string() }),
+  inputSchema: z.object({ city: z.string() }),
   execute: mockWeatherTool,
 });
 
diff --git a/docs/docs/pages/testing-guides/tool-calling.mdx b/docs/docs/pages/testing-guides/tool-calling.mdx
index fd5c4ded9..4bf2138af 100644
--- a/docs/docs/pages/testing-guides/tool-calling.mdx
+++ b/docs/docs/pages/testing-guides/tool-calling.mdx
@@ -309,16 +309,11 @@ const agent = (): AgentAdapter => ({
   call: async (input) => {
     const response = await generateText({
       model: openai("gpt-4.1-nano"),
-      messages: [
-        {
-          role: "system",
-          content: `
+      instructions: `
             You are a helpful assistant that may help the user with weather information.
             Do not guess the city if they don't provide it.
           `,
-        },
-        ...input.messages,
-      ],
+      messages: input.messages,
       tools: { get_current_weather: getCurrentWeather },
       toolChoice: "auto",
     });
diff --git a/javascript/README.md b/javascript/README.md
index 8714e59da..afb6d9b42 100644
--- a/javascript/README.md
+++ b/javascript/README.md
@@ -27,6 +27,17 @@ npm install @langwatch/scenario
 yarn add @langwatch/scenario
 ```
 
+Scenario targets [AI SDK](https://ai-sdk.dev) 7 (`ai@^7`) and Node.js 22.12 or later. If your project is still on AI SDK 6, stay on `@langwatch/scenario` 1.x until you upgrade.
+
+AI SDK 7 emits no spans until a telemetry integration is registered. To let the judge and LangWatch see your agent's LLM calls, register one once at start-up (for example in a Vitest setup file):
+
+```typescript
+import { OpenTelemetry } from "@ai-sdk/otel";
+import { registerTelemetry } from "ai";
+
+registerTelemetry(new OpenTelemetry());
+```
+
 ## Quick Start
 
 Create your first scenario test in under a minute.
@@ -86,7 +97,7 @@ describe("Weather Agent", () => {
     // 1. Define the tools your agent can use
     const getCurrentWeather = tool({
       description: "Get the current weather in a given city.",
-      parameters: z.object({
+      inputSchema: z.object({
         city: z.string().describe("The city to get the weather for."),
       }),
       execute: async ({ city }) =>
@@ -100,7 +111,7 @@ describe("Weather Agent", () => {
       call: async (input) => {
         const response = await generateText({
           model: openai("gpt-4.1-mini"),
-          system: `You are a helpful assistant that may help the user with weather information.`,
+          instructions: `You are a helpful assistant that may help the user with weather information.`,
           messages: input.messages,
           tools: { get_current_weather: getCurrentWeather },
         });
@@ -113,6 +124,7 @@ describe("Weather Agent", () => {
             {
               toolCallId: toolCall.toolCallId,
               messages: input.messages,
+              context: {},
             }
           );
           return [
diff --git a/javascript/examples/vitest/package.json b/javascript/examples/vitest/package.json
index 2b2737a67..d90555d02 100644
--- a/javascript/examples/vitest/package.json
+++ b/javascript/examples/vitest/package.json
@@ -21,11 +21,12 @@
     "vitest-mock-extended": "5.1.1"
   },
   "dependencies": {
-    "@ai-sdk/anthropic": "3.0.41",
-    "@ai-sdk/xai": "3.0.50",
+    "@ai-sdk/anthropic": "4.0.79",
+    "@ai-sdk/otel": "^1.0.130",
+    "@ai-sdk/xai": "5.0.21",
     "@langwatch/scenario": "workspace:*",
     "@openai/agents": "^0.16.0",
-    "ai": "^6.0.0",
+    "ai": "^7.0.130",
     "express": "^5.1.0",
     "openai": "^6.9.1",
     "vite": "^8.2.1",
diff --git a/javascript/examples/vitest/tests/00-demo-light.test.ts b/javascript/examples/vitest/tests/00-demo-light.test.ts
index df0857c6c..07811b718 100644
--- a/javascript/examples/vitest/tests/00-demo-light.test.ts
+++ b/javascript/examples/vitest/tests/00-demo-light.test.ts
@@ -25,15 +25,10 @@ describe("Demo: Lightweight Scenarios", () => {
     call: async (input) => {
       const response = await generateText({
         model: openai("gpt-5-mini"),
-        messages: [
-          {
-            role: "system",
-            content:
-              "You can analyze images and respond briefly and helpfully.",
-          },
-          ...input.messages,
-        ],
-        experimental_telemetry: { isEnabled: true },
+        instructions:
+          "You can analyze images and respond briefly and helpfully.",
+        messages: input.messages,
+        telemetry: { isEnabled: true },
       });
       return response.text;
     },
diff --git a/javascript/examples/vitest/tests/api-service-mocking.test.ts b/javascript/examples/vitest/tests/api-service-mocking.test.ts
index 76b9e29bd..aea42f2c0 100644
--- a/javascript/examples/vitest/tests/api-service-mocking.test.ts
+++ b/javascript/examples/vitest/tests/api-service-mocking.test.ts
@@ -48,7 +48,7 @@ const userDataAgent: AgentAdapter = {
       messages: input.messages,
       tools: { fetch_user_data: fetchUserDataTool },
       toolChoice: "auto",
-      experimental_telemetry: { isEnabled: true },
+      telemetry: { isEnabled: true },
     });
     return response.text;
   },
diff --git a/javascript/examples/vitest/tests/database-tool-mocking.test.ts b/javascript/examples/vitest/tests/database-tool-mocking.test.ts
index 29e8b4a75..bbcefc800 100644
--- a/javascript/examples/vitest/tests/database-tool-mocking.test.ts
+++ b/javascript/examples/vitest/tests/database-tool-mocking.test.ts
@@ -52,7 +52,7 @@ const databaseAgent: AgentAdapter = {
         find_user: findUserTool,
       },
       toolChoice: "auto",
-      experimental_telemetry: { isEnabled: true },
+      telemetry: { isEnabled: true },
     });
     return response.text;
   },
diff --git a/javascript/examples/vitest/tests/false-assumptions.test.ts b/javascript/examples/vitest/tests/false-assumptions.test.ts
index 1921482ee..9e14065c9 100644
--- a/javascript/examples/vitest/tests/false-assumptions.test.ts
+++ b/javascript/examples/vitest/tests/false-assumptions.test.ts
@@ -16,11 +16,9 @@ describe("False Assumptions", () => {
       call: async (input) => {
         const response = await generateText({
           model: openai("gpt-5-mini"),
-          messages: [
-            { role: "system", content: "You are a helpful assistant" },
-            ...input.messages,
-          ],
-          experimental_telemetry: { isEnabled: true },
+          instructions: "You are a helpful assistant",
+          messages: input.messages,
+          telemetry: { isEnabled: true },
         });
 
         return response.text;
diff --git a/javascript/examples/vitest/tests/helpers/register-ai-telemetry.ts b/javascript/examples/vitest/tests/helpers/register-ai-telemetry.ts
new file mode 100644
index 000000000..f29de13b5
--- /dev/null
+++ b/javascript/examples/vitest/tests/helpers/register-ai-telemetry.ts
@@ -0,0 +1,8 @@
+/**
+ * AI SDK 7 emits no spans until a telemetry integration is registered.
+ * Loaded once per worker as a vitest setup file.
+ */
+import { OpenTelemetry } from "@ai-sdk/otel";
+import { registerTelemetry } from "ai";
+
+registerTelemetry(new OpenTelemetry());
diff --git a/javascript/examples/vitest/tests/judge-tool-calls-without-telemetry.test.ts b/javascript/examples/vitest/tests/judge-tool-calls-without-telemetry.test.ts
index e818c0cd5..8da890d09 100644
--- a/javascript/examples/vitest/tests/judge-tool-calls-without-telemetry.test.ts
+++ b/javascript/examples/vitest/tests/judge-tool-calls-without-telemetry.test.ts
@@ -42,7 +42,7 @@ const weatherAgent: AgentAdapter = {
       messages: input.messages,
       tools: { get_weather: getWeatherTool },
       toolChoice: "auto",
-      experimental_telemetry: { isEnabled: true },
+      telemetry: { isEnabled: true },
     });
 
     if (response.toolCalls && response.toolCalls.length > 0) {
diff --git a/javascript/examples/vitest/tests/llm-provider-mocking.test.ts b/javascript/examples/vitest/tests/llm-provider-mocking.test.ts
index ebbf30132..7d2aaecc3 100644
--- a/javascript/examples/vitest/tests/llm-provider-mocking.test.ts
+++ b/javascript/examples/vitest/tests/llm-provider-mocking.test.ts
@@ -1,6 +1,6 @@
 import { openai } from "@ai-sdk/openai";
 import scenario, { type AgentAdapter, AgentRole } from "@langwatch/scenario";
-import { generateText, ToolSet, GenerateTextResult } from "ai";
+import { generateText } from "ai";
 import { describe, it, expect, vi } from "vitest";
 
 /**
@@ -39,7 +39,7 @@ const chatAgent: AgentAdapter = {
     const response = await generateText({
       model: openai("gpt-5-mini"),
       messages: input.messages,
-      experimental_telemetry: { isEnabled: true },
+      telemetry: { isEnabled: true },
     });
     return response.text;
   },
@@ -51,7 +51,7 @@ describe("LLM Provider Mocking", () => {
     // No actual LLM call will be made
     mockGenerateText.mockResolvedValue({
       text: "I can help you with that request.",
-    } as GenerateTextResult<ToolSet, never>);
+    } as Awaited<ReturnType<typeof generateText>>);
 
     const result = await scenario.run({
       name: "llm mock test",
diff --git a/javascript/examples/vitest/tests/mocked-weather-agent-tool.test.ts b/javascript/examples/vitest/tests/mocked-weather-agent-tool.test.ts
index ba6a7312f..1f4594778 100644
--- a/javascript/examples/vitest/tests/mocked-weather-agent-tool.test.ts
+++ b/javascript/examples/vitest/tests/mocked-weather-agent-tool.test.ts
@@ -26,19 +26,14 @@ describe("Mocked Weather Agent Tool", () => {
       call: async (input) => {
         const response = await generateText({
           model: openai("gpt-5-mini"),
-          messages: [
-            {
-              role: "system",
-              content: `
+          instructions: `
                 You are a helpful assistant that may help the user with weather information.
                 Do not guess the city if they don't provide it.
               `,
-            },
-            ...input.messages,
-          ],
+          messages: input.messages,
           tools: { get_current_weather: getCurrentWeather },
           toolChoice: "auto",
-          experimental_telemetry: { isEnabled: true },
+          telemetry: { isEnabled: true },
         });
 
         return response.text;
diff --git a/javascript/examples/vitest/tests/multilingual-agent.test.ts b/javascript/examples/vitest/tests/multilingual-agent.test.ts
index 0cbc0e361..89b58f879 100644
--- a/javascript/examples/vitest/tests/multilingual-agent.test.ts
+++ b/javascript/examples/vitest/tests/multilingual-agent.test.ts
@@ -16,7 +16,7 @@ const createMultilingualAgent = (): AgentAdapter => ({
     const response = await generateText({
       model: openai("gpt-5-mini"),
       messages: input.messages,
-      experimental_telemetry: { isEnabled: true },
+      telemetry: { isEnabled: true },
     });
 
     return response.text;
diff --git a/javascript/examples/vitest/tests/multimodal-images.test.ts b/javascript/examples/vitest/tests/multimodal-images.test.ts
index 27f7f1f29..693e2e7dc 100644
--- a/javascript/examples/vitest/tests/multimodal-images.test.ts
+++ b/javascript/examples/vitest/tests/multimodal-images.test.ts
@@ -35,18 +35,13 @@ describe("Multimodal Images Tests", () => {
     call: async (input) => {
       const response = await generateText({
         model: openai("gpt-5-mini"),
-        messages: [
-          {
-            role: "system",
-            content: `
+        instructions: `
               You are a helpful assistant that can process both text and image input.
               When analyzing images, be descriptive and helpful.
               Respond naturally to user queries about images.
             `,
-          },
-          ...input.messages,
-        ],
-        experimental_telemetry: { isEnabled: true },
+        messages: input.messages,
+        telemetry: { isEnabled: true },
       });
       return response.text;
     },
diff --git a/javascript/examples/vitest/tests/running-in-parallel.test.ts b/javascript/examples/vitest/tests/running-in-parallel.test.ts
index b5b67fb27..9c392496e 100644
--- a/javascript/examples/vitest/tests/running-in-parallel.test.ts
+++ b/javascript/examples/vitest/tests/running-in-parallel.test.ts
@@ -9,18 +9,13 @@ describe("Vegetarian Recipe Agent (Parallel)", () => {
     call: async (input) => {
       const response = await generateText({
         model: openai("gpt-5-mini"),
-        messages: [
-          {
-            role: "system",
-            content: `You are a vegetarian recipe agent. When a user asks for a recipe,
+        instructions: `You are a vegetarian recipe agent. When a user asks for a recipe,
             <rules>
               - You may ask ONLY ONE clarifying question if absolutely necessary (like dietary restrictions or cooking time preference), but then you MUST provide a complete vegetarian recipe.
               - Do not ask multiple follow-up questions.
             </rules>`,
-          },
-          ...input.messages,
-        ],
-        experimental_telemetry: { isEnabled: true },
+        messages: input.messages,
+        telemetry: { isEnabled: true },
       });
       return response.text;
     },
diff --git a/javascript/examples/vitest/tests/simple-tool-mocking.test.ts b/javascript/examples/vitest/tests/simple-tool-mocking.test.ts
index 2f4907a2f..c9d54a3ea 100644
--- a/javascript/examples/vitest/tests/simple-tool-mocking.test.ts
+++ b/javascript/examples/vitest/tests/simple-tool-mocking.test.ts
@@ -39,7 +39,7 @@ const userDataAgent: AgentAdapter = {
       messages: input.messages,
       tools: { fetch_user_data: fetchUserDataTool },
       toolChoice: "auto",
-      experimental_telemetry: { isEnabled: true },
+      telemetry: { isEnabled: true },
     });
     return response.text;
   },
diff --git a/javascript/examples/vitest/tests/span-based-evaluation.test.ts b/javascript/examples/vitest/tests/span-based-evaluation.test.ts
index a59dcade9..7bd4c1a08 100644
--- a/javascript/examples/vitest/tests/span-based-evaluation.test.ts
+++ b/javascript/examples/vitest/tests/span-based-evaluation.test.ts
@@ -1,7 +1,7 @@
 import { openai } from "@ai-sdk/openai";
 import scenario, { type AgentAdapter, AgentRole } from "@langwatch/scenario";
 import { trace, context } from "@opentelemetry/api";
-import { generateText, stepCountIs, tool } from "ai";
+import { generateText, isStepCount, tool } from "ai";
 import { describe, it, expect } from "vitest";
 import { z } from "zod/v4";
 
@@ -68,17 +68,12 @@ const observableAgent: AgentAdapter = {
       // LLM call with tool usage
       const response = await generateText({
         model: openai("gpt-5-mini"),
-        messages: [
-          {
-            role: "system",
-            content: `You are an order processing assistant.
+        instructions: `You are an order processing assistant.
 When asked about products, use the check_inventory tool.`,
-          },
-          ...input.messages,
-        ],
+        messages: input.messages,
         tools: { check_inventory: checkInventoryTool },
-        experimental_telemetry: { isEnabled: true },
-        stopWhen: stepCountIs(2),
+        telemetry: { isEnabled: true },
+        stopWhen: isStepCount(2),
       });
 
       return response.text;
diff --git a/javascript/examples/vitest/tests/testing-remote-agents-json.test.ts b/javascript/examples/vitest/tests/testing-remote-agents-json.test.ts
index 010710ff9..24e923fc7 100644
--- a/javascript/examples/vitest/tests/testing-remote-agents-json.test.ts
+++ b/javascript/examples/vitest/tests/testing-remote-agents-json.test.ts
@@ -31,21 +31,18 @@ beforeAll(async () => {
           // Use real LLM to generate response
           const result = await generateText({
             model: openai("gpt-5-mini"),
+            /**
+             * Simple system prompt to make the agent pretend like it has access to a weather API and make up the weather so the tests pass.
+             */
+            instructions:
+              "You are a helpful weather assistant. Provide brief, friendly responses about weather. Pretend like you have access to a weather API and make up the weather.",
             messages: [
-              {
-                role: "system",
-                /**
-                 * Simple system prompt to make the agent pretend like it has access to a weather API and make up the weather so the tests pass.
-                 */
-                content:
-                  "You are a helpful weather assistant. Provide brief, friendly responses about weather. Pretend like you have access to a weather API and make up the weather.",
-              },
               {
                 role: "user",
                 content: message,
               },
             ],
-            experimental_telemetry: { isEnabled: true },
+            telemetry: { isEnabled: true },
           });
 
           res.writeHead(200, { "Content-Type": "application/json" });
diff --git a/javascript/examples/vitest/tests/testing-remote-agents-sse.test.ts b/javascript/examples/vitest/tests/testing-remote-agents-sse.test.ts
index 505c406db..be84eb6a6 100644
--- a/javascript/examples/vitest/tests/testing-remote-agents-sse.test.ts
+++ b/javascript/examples/vitest/tests/testing-remote-agents-sse.test.ts
@@ -44,18 +44,15 @@ beforeAll(async () => {
           // Stream response using real LLM
           const result = streamText({
             model: openai("gpt-5-mini"),
+            instructions:
+              "You are a helpful weather assistant. Provide brief, friendly responses. Pretend you have access to weather data. Pretend like you have access to a weather API and make up the weather.",
             messages: [
-              {
-                role: "system",
-                content:
-                  "You are a helpful weather assistant. Provide brief, friendly responses. Pretend you have access to weather data. Pretend like you have access to a weather API and make up the weather.",
-              },
               {
                 role: "user",
                 content,
               },
             ],
-            experimental_telemetry: { isEnabled: true },
+            telemetry: { isEnabled: true },
           });
 
           // Stream chunks in SSE format
diff --git a/javascript/examples/vitest/tests/testing-remote-agents-stateful.test.ts b/javascript/examples/vitest/tests/testing-remote-agents-stateful.test.ts
index ee440c05d..eda2e0dad 100644
--- a/javascript/examples/vitest/tests/testing-remote-agents-stateful.test.ts
+++ b/javascript/examples/vitest/tests/testing-remote-agents-stateful.test.ts
@@ -42,15 +42,10 @@ beforeAll(async () => {
           // Generate response with full history
           const result = await generateText({
             model: openai("gpt-5-mini"),
-            messages: [
-              {
-                role: "system",
-                content:
-                  "You are a helpful travel assistant. Help users plan their trips and provide information about destinations.",
-              },
-              ...history,
-            ],
-            experimental_telemetry: { isEnabled: true },
+            instructions:
+              "You are a helpful travel assistant. Help users plan their trips and provide information about destinations.",
+            messages: history,
+            telemetry: { isEnabled: true },
           });
 
           // Add assistant response to history
diff --git a/javascript/examples/vitest/tests/testing-remote-agents-streaming.test.ts b/javascript/examples/vitest/tests/testing-remote-agents-streaming.test.ts
index 0988e0994..caa2c33d1 100644
--- a/javascript/examples/vitest/tests/testing-remote-agents-streaming.test.ts
+++ b/javascript/examples/vitest/tests/testing-remote-agents-streaming.test.ts
@@ -42,18 +42,15 @@ beforeAll(async () => {
 
           const result = streamText({
             model: openai("gpt-5-mini"),
+            instructions:
+              "You are a helpful weather assistant. Provide brief, friendly responses, immediately. Pretend like you have access to a weather API and make up the weather.",
             messages: [
-              {
-                role: "system",
-                content:
-                  "You are a helpful weather assistant. Provide brief, friendly responses, immediately. Pretend like you have access to a weather API and make up the weather.",
-              },
               {
                 role: "user",
                 content,
               },
             ],
-            experimental_telemetry: { isEnabled: true },
+            telemetry: { isEnabled: true },
           });
 
           // Stream chunks to client
diff --git a/javascript/examples/vitest/tests/testing-remote-agents-traces.test.ts b/javascript/examples/vitest/tests/testing-remote-agents-traces.test.ts
index 12b2d0a4e..bdc4eb2b8 100644
--- a/javascript/examples/vitest/tests/testing-remote-agents-traces.test.ts
+++ b/javascript/examples/vitest/tests/testing-remote-agents-traces.test.ts
@@ -3,7 +3,7 @@ import { openai } from "@ai-sdk/openai";
 import scenario, { AgentRole, type AgentAdapter } from "@langwatch/scenario";
 import { context, propagation } from "@opentelemetry/api";
 import { W3CTraceContextPropagator } from "@opentelemetry/core";
-import { generateText, ModelMessage, stepCountIs, tool } from "ai";
+import { generateText, ModelMessage, isStepCount, tool } from "ai";
 import { describe, it, expect, beforeAll, afterAll } from "vitest";
 import { z } from "zod/v4";
 
@@ -60,17 +60,12 @@ beforeAll(async () => {
           const result = await context.with(extracted, () =>
             generateText({
               model: openai("gpt-5-mini"),
-              messages: [
-                {
-                  role: "system",
-                  content:
-                    "You are an order support assistant. Use the lookup_order_status tool to check orders before answering.",
-                },
-                ...messages,
-              ],
+              instructions:
+                "You are an order support assistant. Use the lookup_order_status tool to check orders before answering.",
+              messages: messages,
               tools: { lookup_order_status: lookupOrderStatus },
-              stopWhen: stepCountIs(3),
-              experimental_telemetry: { isEnabled: true },
+              stopWhen: isStepCount(3),
+              telemetry: { isEnabled: true },
             })
           );
 
diff --git a/javascript/examples/vitest/tests/tool-call-role-reversal.test.ts b/javascript/examples/vitest/tests/tool-call-role-reversal.test.ts
index 1d7fe42e6..9a82fdbd7 100644
--- a/javascript/examples/vitest/tests/tool-call-role-reversal.test.ts
+++ b/javascript/examples/vitest/tests/tool-call-role-reversal.test.ts
@@ -26,23 +26,19 @@ const toolCallingAgent: AgentAdapter = {
   call: async (input) => {
     const response = await generateText({
       model: openai("gpt-5-mini"),
-      messages: [
-        {
-          role: "system",
-          content: "You are a helpful shopping assistant. Always use the lookup tool to check product details before responding.",
-        },
-        ...input.messages,
-      ],
+      instructions:
+        "You are a helpful shopping assistant. Always use the lookup tool to check product details before responding.",
+      messages: input.messages,
       tools: { lookup_product: lookupTool },
       toolChoice: "auto",
-      experimental_telemetry: { isEnabled: true },
+      telemetry: { isEnabled: true },
     });
 
     if (response.toolCalls && response.toolCalls.length > 0) {
       const toolCall = response.toolCalls[0];
       const toolResult = await lookupTool.execute!(
         toolCall.input as { product: string },
-        { toolCallId: toolCall.toolCallId, messages: input.messages }
+        { toolCallId: toolCall.toolCallId, messages: input.messages, context: {} }
       );
       return [
         {
diff --git a/javascript/examples/vitest/tests/tool-failure-simulation.test.ts b/javascript/examples/vitest/tests/tool-failure-simulation.test.ts
index 8f7c2587e..af1848848 100644
--- a/javascript/examples/vitest/tests/tool-failure-simulation.test.ts
+++ b/javascript/examples/vitest/tests/tool-failure-simulation.test.ts
@@ -45,7 +45,7 @@ const resilientAgent: AgentAdapter = {
       messages: input.messages,
       tools: { call_external_service: callExternalServiceTool },
       toolChoice: "auto",
-      experimental_telemetry: { isEnabled: true },
+      telemetry: { isEnabled: true },
     });
 
     // Check if the LLM decided to call any tools
diff --git a/javascript/examples/vitest/tests/vegetarian-recipe-agent.test.ts b/javascript/examples/vitest/tests/vegetarian-recipe-agent.test.ts
index b8e4756d4..f1ea44970 100644
--- a/javascript/examples/vitest/tests/vegetarian-recipe-agent.test.ts
+++ b/javascript/examples/vitest/tests/vegetarian-recipe-agent.test.ts
@@ -9,19 +9,14 @@ describe("Vegetarian Recipe Agent", () => {
     call: async (input) => {
       const response = await generateText({
         model: openai("gpt-5-mini"),
-        messages: [
-          {
-            role: "system",
-            content: `
+        instructions: `
               You are a vegetarian recipe agent.
               Given the user request, ask AT MOST ONE follow-up question,
               then provide a complete recipe. Once you have given a recipe, answer
               without asking further questions. Keep your responses concise and focused.
             `,
-          },
-          ...input.messages,
-        ],
-        experimental_telemetry: { isEnabled: true },
+        messages: input.messages,
+        telemetry: { isEnabled: true },
       });
       return response.text;
     },
diff --git a/javascript/examples/vitest/tests/weather-agent.test.ts b/javascript/examples/vitest/tests/weather-agent.test.ts
index dab160019..f3334d2e4 100644
--- a/javascript/examples/vitest/tests/weather-agent.test.ts
+++ b/javascript/examples/vitest/tests/weather-agent.test.ts
@@ -24,19 +24,14 @@ const weatherAgent: AgentAdapter = {
   call: async (input) => {
     const response = await generateText({
       model: openai("gpt-5-mini"),
-      messages: [
-        {
-          role: "system",
-          content: `
+      instructions: `
             You are a helpful assistant that may help the user with weather information.
             Do not guess the city if they don't provide it.
           `,
-        },
-        ...input.messages,
-      ],
+      messages: input.messages,
       tools: { get_current_weather: getCurrentWeather },
       toolChoice: "auto",
-      experimental_telemetry: { isEnabled: true },
+      telemetry: { isEnabled: true },
     });
 
     if (response.toolCalls && response.toolCalls.length > 0) {
@@ -47,6 +42,7 @@ const weatherAgent: AgentAdapter = {
         {
           toolCallId: toolCall.toolCallId,
           messages: input.messages,
+          context: {},
         }
       );
       return [
diff --git a/javascript/examples/vitest/vitest.config.ts b/javascript/examples/vitest/vitest.config.ts
index ff8f9830d..6034fa312 100644
--- a/javascript/examples/vitest/vitest.config.ts
+++ b/javascript/examples/vitest/vitest.config.ts
@@ -39,8 +39,10 @@ export default withScenario(
       testTimeout: 180_000, // 3 minutes
 
       // Load environment variables from .env file
-      // Required for API keys and configuration needed by AI agents
-      setupFiles: ["dotenv/config"],
+      // Required for API keys and configuration needed by AI agents.
+      // Then register the AI SDK 7 OpenTelemetry integration so agent
+      // calls emit spans the judge and LangWatch can see.
+      setupFiles: ["dotenv/config", "./tests/helpers/register-ai-telemetry.ts"],
 
       // (Optional) Use Scenario's custom reporter for better test output
       // Provides detailed information about agent interactions,
diff --git a/javascript/package.json b/javascript/package.json
index 19a252bdf..9b26a2b1f 100644
--- a/javascript/package.json
+++ b/javascript/package.json
@@ -50,24 +50,24 @@
   "author": "LangWatch",
   "license": "Apache-2.0",
   "engines": {
-    "node": ">=20",
+    "node": ">=22.12",
     "pnpm": ">=8"
   },
   "packageManager": "pnpm@10.22.0",
   "dependencies": {
     "@ag-ui/core": "^0.0.58",
-    "@ai-sdk/openai": "^3.0.26",
+    "@ai-sdk/openai": "^4.0.86",
     "@openai/agents": "^0.16.0",
     "@opentelemetry/api": "^1.9.0",
     "@opentelemetry/context-async-hooks": "2.10.0",
     "@opentelemetry/sdk-node": "0.221.0",
     "@opentelemetry/sdk-trace-base": "2.10.0",
     "@opentelemetry/sdk-trace-node": "2.10.0",
-    "ai": "^6.0.0",
+    "ai": "^7.0.130",
     "chalk": "^5.6.2",
     "ffmpeg-static": "5.3.0",
     "fft.js": "4.0.4",
-    "langwatch": ">=0.16.1 <2.0.0",
+    "langwatch": "^1.20.0",
     "open": "11.0.1",
     "openai": "^6.9.1",
     "rxjs": "^7.8.2",
@@ -76,8 +76,8 @@
     "zod": "^3.25.76"
   },
   "devDependencies": {
-    "@ai-sdk/amazon-bedrock": "^4.0.183",
-    "@ai-sdk/google": "^3.0.127",
+    "@ai-sdk/amazon-bedrock": "^5.0.115",
+    "@ai-sdk/google": "^4.0.95",
     "@amiceli/vitest-cucumber": "^7.0.0",
     "@elevenlabs/elevenlabs-js": "^2.51.0",
     "@eslint/js": "^9.39.1",
@@ -132,7 +132,7 @@
   "peerDependencies": {
     "@elevenlabs/elevenlabs-js": ">=2.51.0",
     "@google/genai": ">=2.0.0",
-    "ai": ">=6.0.0",
+    "ai": ">=7.0.0",
     "vitest": ">=3.2.4"
   },
   "peerDependenciesMeta": {
diff --git a/javascript/src/agents/__tests__/realtime-echo-safe.test.ts b/javascript/src/agents/__tests__/realtime-echo-safe.test.ts
index 06c5bb135..b406d9d26 100644
--- a/javascript/src/agents/__tests__/realtime-echo-safe.test.ts
+++ b/javascript/src/agents/__tests__/realtime-echo-safe.test.ts
@@ -177,11 +177,11 @@ describe("realtime echo-safety (AC-JS1')", () => {
     // red-capable at the JS layer: the realtime turn goes straight through the
     // production reversal, and the parrot sees the agent's transcript verbatim.
     const naiveReversed = messageRoleReversal([
-      { role: "system", content: "you are pretending to be a user" },
       { role: "assistant", content: "Hello, how can I help you today" },
       agentTurn,
     ]);
     const naiveResult = await parrotInvokeLLM({
+      instructions: "you are pretending to be a user",
       messages: naiveReversed,
     } as unknown as InvokeLLMParams);
     const naiveAnswer = naiveResult.text ?? "";
diff --git a/javascript/src/agents/__tests__/user-simulator-instructions.test.ts b/javascript/src/agents/__tests__/user-simulator-instructions.test.ts
new file mode 100644
index 000000000..65eb18d2f
--- /dev/null
+++ b/javascript/src/agents/__tests__/user-simulator-instructions.test.ts
@@ -0,0 +1,90 @@
+/**
+ * The user simulator sends its system prompt as AI SDK 7 `instructions`.
+ *
+ * Covers the user simulator @unit scenarios of
+ * `specs/ai-sdk-7-system-prompts.feature`. AI SDK 7 refuses system messages
+ * inside `messages` unless the call allows them.
+ */
+import type { ModelMessage } from "ai";
+import { describe, it, expect, vi } from "vitest";
+
+import type { AgentInput } from "../../domain";
+import type { InvokeLLMParams } from "../types";
+import { userSimulatorAgent } from "../user-simulator-agent";
+
+vi.mock("../../config", () => ({
+  getProjectConfig: vi.fn().mockResolvedValue({
+    defaultModel: { model: "openai/gpt-5-mini", temperature: 0 },
+  }),
+}));
+
+function makeInput(messages: ModelMessage[]): AgentInput {
+  return {
+    threadId: "t-instructions",
+    messages,
+    newMessages: [],
+    requestedRole: "User" as AgentInput["requestedRole"],
+    scenarioConfig: {
+      name: "test",
+      description: "A test scenario description",
+    } as unknown as AgentInput["scenarioConfig"],
+    scenarioState: {} as AgentInput["scenarioState"],
+  };
+}
+
+/** Stub the simulator's LLM and capture the parameters of its last call. */
+function captureLLM(simulator: ReturnType<typeof userSimulatorAgent>) {
+  const seen: { params?: InvokeLLMParams } = {};
+  (
+    simulator as unknown as {
+      invokeLLM: (
+        params: InvokeLLMParams
+      ) => Promise<{ text: string; toolCalls: []; steps: [] }>;
+    }
+  ).invokeLLM = async (params) => {
+    seen.params = params;
+    return { text: "Where is my order?", toolCalls: [], steps: [] };
+  };
+  return seen;
+}
+
+describe("given a user simulator agent whose model is stubbed", () => {
+  describe("when the simulator produces the user's turn", () => {
+    /** @scenario "The user simulator sends its system prompt as instructions" */
+    it("carries the system prompt as instructions, not as a message", async () => {
+      const simulator = userSimulatorAgent({ systemPrompt: "Be a customer" });
+      const seen = captureLLM(simulator);
+
+      await simulator.call(
+        makeInput([{ role: "assistant", content: "Hi, how can I help?" }])
+      );
+
+      expect(seen.params?.instructions).toBe("Be a customer");
+      expect(seen.params?.messages?.some((m) => m.role === "system")).toBe(
+        false
+      );
+    });
+  });
+
+  describe("when the transcript holds a system turn added by the script", () => {
+    /** @scenario "A system turn added by the script still reaches the user simulator" */
+    it("allows system messages and keeps the scripted turn", async () => {
+      const simulator = userSimulatorAgent({ systemPrompt: "Be a customer" });
+      const seen = captureLLM(simulator);
+      const scripted: ModelMessage = {
+        role: "system",
+        content: "The user is now in a hurry",
+      };
+
+      await simulator.call(
+        makeInput([
+          { role: "assistant", content: "Hi, how can I help?" },
+          scripted,
+        ])
+      );
+
+      expect(seen.params?.allowSystemInMessages).toBe(true);
+      expect(seen.params?.messages).toContainEqual(scripted);
+    });
+  });
+});
diff --git a/javascript/src/agents/__tests__/user-simulator-voice.test.ts b/javascript/src/agents/__tests__/user-simulator-voice.test.ts
index c1cdb6bb5..a07b09132 100644
--- a/javascript/src/agents/__tests__/user-simulator-voice.test.ts
+++ b/javascript/src/agents/__tests__/user-simulator-voice.test.ts
@@ -271,11 +271,10 @@ describeFeature(
             // Capture system prompt via LLM spy — check persona block present.
             systemPromptCapture = "";
             (sim as unknown as {
-              invokeLLM: (p: { messages: Array<{ role: string; content: string }> }) => Promise<{ text: string; toolCalls: []; steps: [] }>;
-            }).invokeLLM = async ({ messages }) => {
-              // System message is passed unchanged through role reversal.
-              const sysMsg = messages.find((m) => m.role === "system");
-              systemPromptCapture = sysMsg?.content ?? "";
+              invokeLLM: (p: { instructions?: string }) => Promise<{ text: string; toolCalls: []; steps: [] }>;
+            }).invokeLLM = async ({ instructions }) => {
+              // The system prompt rides on `instructions`, outside the messages.
+              systemPromptCapture = instructions ?? "";
               return { text: "I need help please", toolCalls: [], steps: [] };
             };
 
diff --git a/javascript/src/agents/judge/__tests__/judge-agent.test.ts b/javascript/src/agents/judge/__tests__/judge-agent.test.ts
index e4828d63c..3177a7723 100644
--- a/javascript/src/agents/judge/__tests__/judge-agent.test.ts
+++ b/javascript/src/agents/judge/__tests__/judge-agent.test.ts
@@ -218,6 +218,7 @@ describe("JudgeAgent", () => {
   });
 
   describe("when custom system prompt is used with large trace", () => {
+    /** @scenario "The judge sends its system prompt as instructions" */
     it("preserves the custom system prompt", async () => {
       const largeTrace = createLargeTrace();
       const collector = createMockSpanCollector(largeTrace);
@@ -244,15 +245,16 @@ describe("JudgeAgent", () => {
 
       await agent.call(createBaseInput());
 
-      const systemMsg = capturedParams!.messages?.find(
-        (m) => "role" in m && m.role === "system"
-      );
-      expect(systemMsg).toBeDefined();
+      const instructions = capturedParams!.instructions;
+      expect(instructions).toBeDefined();
       const content =
-        typeof systemMsg!.content === "string"
-          ? systemMsg!.content
-          : JSON.stringify(systemMsg!.content);
+        typeof instructions === "string"
+          ? instructions
+          : JSON.stringify(instructions);
       expect(content).toContain(customPrompt);
+      expect(
+        capturedParams!.messages?.some((m) => m.role === "system")
+      ).toBe(false);
     });
   });
 
diff --git a/javascript/src/agents/judge/__tests__/judge-per-criterion-verdicts.unit.test.ts b/javascript/src/agents/judge/__tests__/judge-per-criterion-verdicts.unit.test.ts
index 044910592..f5b8e2405 100644
--- a/javascript/src/agents/judge/__tests__/judge-per-criterion-verdicts.unit.test.ts
+++ b/javascript/src/agents/judge/__tests__/judge-per-criterion-verdicts.unit.test.ts
@@ -118,8 +118,7 @@ function finishTestSchema(params: InvokeLLMParams): Record<string, unknown> {
 }
 
 function systemPrompt(params: InvokeLLMParams): string {
-  const first = params.messages![0]!;
-  return typeof first.content === "string" ? first.content : "";
+  return typeof params.instructions === "string" ? params.instructions : "";
 }
 
 class ScriptedAgent extends AgentAdapter {
diff --git a/javascript/src/agents/judge/__tests__/judge-provider-matrix.integration.test.ts b/javascript/src/agents/judge/__tests__/judge-provider-matrix.integration.test.ts
index 7a60aa58f..29d7078c1 100644
--- a/javascript/src/agents/judge/__tests__/judge-provider-matrix.integration.test.ts
+++ b/javascript/src/agents/judge/__tests__/judge-provider-matrix.integration.test.ts
@@ -12,7 +12,7 @@ import { dirname, resolve } from "node:path";
 import { fileURLToPath } from "node:url";
 
 import { createAmazonBedrock } from "@ai-sdk/amazon-bedrock";
-import { createGoogleGenerativeAI } from "@ai-sdk/google";
+import { createGoogle } from "@ai-sdk/google";
 import { createOpenAI } from "@ai-sdk/openai";
 import { loadFeature, describeFeature } from "@amiceli/vitest-cucumber";
 import type { LanguageModel, ModelMessage } from "ai";
@@ -58,7 +58,7 @@ const MODELS: Record<string, () => LanguageModel | null> = {
     env.OPENAI_API_KEY ? createOpenAI()("gpt-5.6-luna") : null,
   "gemini gemini-3.5-flash": () =>
     env.GEMINI_API_KEY
-      ? createGoogleGenerativeAI({ apiKey: env.GEMINI_API_KEY })(
+      ? createGoogle({ apiKey: env.GEMINI_API_KEY })(
           "gemini-3.5-flash"
         )
       : null,
diff --git a/javascript/src/agents/judge/__tests__/judge-two-phase.unit.test.ts b/javascript/src/agents/judge/__tests__/judge-two-phase.unit.test.ts
index 72dad8bb2..422c26efd 100644
--- a/javascript/src/agents/judge/__tests__/judge-two-phase.unit.test.ts
+++ b/javascript/src/agents/judge/__tests__/judge-two-phase.unit.test.ts
@@ -167,12 +167,7 @@ function schemaShapeKeys(tool: unknown): string[] {
 }
 
 function systemPromptOf(params: InvokeLLMParams | undefined): string {
-  const systemMessage = params?.messages?.find(
-    (m) => "role" in m && m.role === "system"
-  );
-  return typeof systemMessage?.content === "string"
-    ? systemMessage.content
-    : "";
+  return typeof params?.instructions === "string" ? params.instructions : "";
 }
 
 const feature = await loadFeature(FEATURE_PATH);
diff --git a/javascript/src/agents/judge/__tests__/remote-traces.unit.test.ts b/javascript/src/agents/judge/__tests__/remote-traces.unit.test.ts
index bea4d48f1..45e2b74ee 100644
--- a/javascript/src/agents/judge/__tests__/remote-traces.unit.test.ts
+++ b/javascript/src/agents/judge/__tests__/remote-traces.unit.test.ts
@@ -236,13 +236,13 @@ function userMessageContent(params: InvokeLLMParams | undefined): string {
 }
 
 function systemMessageContent(params: InvokeLLMParams | undefined): string {
-  const systemMessage = params?.messages?.find(
-    (m) => "role" in m && m.role === "system"
-  );
-  if (!systemMessage) throw new Error("Expected a system message in LLM call");
-  return typeof systemMessage.content === "string"
-    ? systemMessage.content
-    : JSON.stringify(systemMessage.content);
+  const instructions = params?.instructions;
+  if (instructions === undefined) {
+    throw new Error("Expected instructions in LLM call");
+  }
+  return typeof instructions === "string"
+    ? instructions
+    : JSON.stringify(instructions);
 }
 
 // ---------------------------------------------------------------------------
diff --git a/javascript/src/agents/judge/judge-agent.ts b/javascript/src/agents/judge/judge-agent.ts
index 5cf6d5b35..a75a4c2a2 100644
--- a/javascript/src/agents/judge/judge-agent.ts
+++ b/javascript/src/agents/judge/judge-agent.ts
@@ -4,7 +4,7 @@ import {
   ToolSet,
   Tool,
   tool,
-  stepCountIs,
+  isStepCount,
   hasToolCall,
 } from "ai";
 import { z } from "zod/v4";
@@ -484,7 +484,7 @@ export class JudgeAgent extends JudgeAgentAdapter {
   /**
    * Extract a string identifier from the configured model for substring matching.
    *
-   * `LanguageModel` is `GlobalProviderModelId | LanguageModelV3 | LanguageModelV2`.
+   * `LanguageModel` is `GlobalProviderModelId | LanguageModelV4 | LanguageModelV3 | LanguageModelV2`.
    * `GlobalProviderModelId` resolves to a string literal; provider objects expose
    * a `modelId` property. We handle both shapes.
    */
@@ -492,7 +492,7 @@ export class JudgeAgent extends JudgeAgentAdapter {
     const model = this.cfg.model;
     if (!model) return "";
     if (typeof model === "string") return model.toLowerCase();
-    // LanguageModelV3 / LanguageModelV2 objects expose `modelId`.
+    // LanguageModelV4 / V3 / V2 objects expose `modelId`.
     const obj = model as { modelId?: string; provider?: string };
     const parts = [obj.provider ?? "", obj.modelId ?? ""].filter(Boolean);
     return parts.join("/").toLowerCase();
@@ -857,10 +857,8 @@ export class JudgeAgent extends JudgeAgentAdapter {
 
     const params: InvokeLLMParams = {
       model: mergedConfig.model,
-      messages: [
-        { role: "system", content: systemPrompt },
-        { role: "user", content: contentForJudge },
-      ],
+      instructions: systemPrompt,
+      messages: [{ role: "user", content: contentForJudge }],
       temperature: mergedConfig.temperature,
       maxOutputTokens: mergedConfig.maxTokens,
       tools,
@@ -868,7 +866,7 @@ export class JudgeAgent extends JudgeAgentAdapter {
     };
     if (isLargeTrace) {
       params.stopWhen = [
-        stepCountIs(this.maxDiscoverySteps),
+        isStepCount(this.maxDiscoverySteps),
         hasToolCall("continue_test"),
         hasToolCall("make_verdict"),
       ];
@@ -895,8 +893,8 @@ export class JudgeAgent extends JudgeAgentAdapter {
       );
       // generateText never writes back into params.messages, so the discovery
       // cycles live only on the completion steps. Recombine both before
-      // collapsing, then drop the two messages the verdict phase rebuilds by
-      // itself (the system prompt and the criteria block).
+      // collapsing, then drop the one message the verdict phase rebuilds by
+      // itself (the criteria block; the system prompt rides on instructions).
       const discoveryHistory: ModelMessage[] = [
         ...(params.messages ?? []),
         ...(completion.steps ?? []).flatMap((step) => [
@@ -905,7 +903,7 @@ export class JudgeAgent extends JudgeAgentAdapter {
       ];
       return {
         decision: "exhausted",
-        discoveryRecap: collapseDiscoveryHistory(discoveryHistory).slice(2),
+        discoveryRecap: collapseDiscoveryHistory(discoveryHistory).slice(1),
       };
     }
     return {
@@ -1044,7 +1042,6 @@ export class JudgeAgent extends JudgeAgentAdapter {
           });
 
       const messages: ModelMessage[] = [
-        { role: "system", content: systemPrompt },
         { role: "user", content: contentForJudge },
       ];
 
@@ -1087,6 +1084,7 @@ export class JudgeAgent extends JudgeAgentAdapter {
 
       const params: InvokeLLMParams = {
         model: mergedConfig.model,
+        instructions: systemPrompt,
         messages,
         temperature: mergedConfig.temperature,
         maxOutputTokens: mergedConfig.maxTokens,
@@ -1098,7 +1096,7 @@ export class JudgeAgent extends JudgeAgentAdapter {
       if (isLargeTrace && !exhaustedEntry) {
         params.toolChoice = "required";
         params.stopWhen = [
-          stepCountIs(this.maxDiscoverySteps),
+          isStepCount(this.maxDiscoverySteps),
           hasToolCall("finish_test"),
           ...(waitExtensionAvailable ? [hasToolCall("wait_for_traces")] : []),
         ];
@@ -1177,10 +1175,10 @@ export class JudgeAgent extends JudgeAgentAdapter {
 
   /**
    * True when `toolName` was called anywhere in the (possibly multi-step)
-   * completion — the aggregate `steps` array when present, else the final
-   * `toolCalls`. AI SDK v6 surfaces only the final step in
-   * `completion.toolCalls`; a terminal call earlier in the loop would be
-   * invisible there.
+   * completion — the aggregate `steps` array when present, else
+   * `toolCalls`. AI SDK 6 surfaced only the final step in
+   * `completion.toolCalls` (AI SDK 7 accumulates every step); `steps` reads
+   * the same under both.
    */
   private completionCalledTool(
     completion: InvokeLLMResult,
diff --git a/javascript/src/agents/llm-invoker.factory.ts b/javascript/src/agents/llm-invoker.factory.ts
index 0d0641b44..bde9c61c3 100644
--- a/javascript/src/agents/llm-invoker.factory.ts
+++ b/javascript/src/agents/llm-invoker.factory.ts
@@ -35,7 +35,7 @@ export const createLLMInvoker = (
       try {
         return await generate({
           ...adapted,
-          experimental_telemetry: { isEnabled: true },
+          telemetry: { isEnabled: true },
         });
       } catch (error) {
         const adaptation = adaptationFor(error, adapted);
diff --git a/javascript/src/agents/red-team/red-team-agent.ts b/javascript/src/agents/red-team/red-team-agent.ts
index 2187381fc..a8b05ce30 100644
--- a/javascript/src/agents/red-team/red-team-agent.ts
+++ b/javascript/src/agents/red-team/red-team-agent.ts
@@ -431,6 +431,9 @@ Reply with exactly this JSON and nothing else:
     const result = await generateText({
       model: this.model,
       messages: this.attackerHistory as ModelMessage[],
+      // The prompt plus [SCORE]/[BACKTRACKED]/[INJECTED] markers are timeline
+      // system turns; on turn 1 the prompt is the only message there is.
+      allowSystemInMessages: true,
       temperature: this.temperature,
       maxOutputTokens: this.maxTokens,
     });
diff --git a/javascript/src/agents/user-simulator-agent.ts b/javascript/src/agents/user-simulator-agent.ts
index ee8f10a62..1831ecccc 100644
--- a/javascript/src/agents/user-simulator-agent.ts
+++ b/javascript/src/agents/user-simulator-agent.ts
@@ -457,7 +457,6 @@ class UserSimulatorAgent extends UserSimulatorAgentAdapter {
     const strippedMessages = stripAudioContent(preparedMessages);
 
     const messages: ModelMessage[] = [
-      { role: "system", content: systemPrompt },
       { role: "assistant", content: "Hello, how can I help you today" },
       ...strippedMessages,
     ];
@@ -486,7 +485,10 @@ class UserSimulatorAgent extends UserSimulatorAgentAdapter {
     ) {
       const completion = await this.invokeLLM({
         model: mergedConfig.model,
+        instructions: systemPrompt,
         messages: reversedMessages,
+        // The transcript may carry system turns a script added with message().
+        allowSystemInMessages: true,
         temperature: mergedConfig.temperature,
         maxOutputTokens: mergedConfig.maxTokens,
       });
diff --git a/javascript/src/voice/__tests__/composable-transcript-attach.test.ts b/javascript/src/voice/__tests__/composable-transcript-attach.test.ts
index 96d2809a7..1dbc06fd7 100644
--- a/javascript/src/voice/__tests__/composable-transcript-attach.test.ts
+++ b/javascript/src/voice/__tests__/composable-transcript-attach.test.ts
@@ -37,13 +37,16 @@ const stubStt: STTProvider = {
 /** Minimal ai-sdk LanguageModel stub returning a fixed reply. */
 function stubLlm(text: string): LanguageModel {
   return {
-    specificationVersion: "v3" as const,
+    specificationVersion: "v4" as const,
     provider: "fake",
     modelId: "fake-model",
     supportedUrls: {},
     doGenerate: async () => ({
-      finishReason: "stop",
-      usage: { inputTokens: 1, outputTokens: 1, totalTokens: 2 },
+      finishReason: { unified: "stop", raw: "stop" },
+      usage: {
+        inputTokens: { total: 1, noCache: 1, cacheRead: 0, cacheWrite: 0 },
+        outputTokens: { total: 1, text: 1, reasoning: 0 },
+      },
       content: [{ type: "text", text }],
       warnings: [],
     }),
diff --git a/javascript/src/voice/__tests__/factories.test.ts b/javascript/src/voice/__tests__/factories.test.ts
index 0313b935a..8fa070116 100644
--- a/javascript/src/voice/__tests__/factories.test.ts
+++ b/javascript/src/voice/__tests__/factories.test.ts
@@ -29,13 +29,16 @@ const stubStt: STTProvider = {
 
 // Minimal ai-sdk LanguageModel stub — never invoked (factory test only).
 const stubLlm = {
-  specificationVersion: "v3" as const,
+  specificationVersion: "v4" as const,
   provider: "fake",
   modelId: "fake-model",
   supportedUrls: {},
   doGenerate: async () => ({
-    finishReason: "stop",
-    usage: { inputTokens: 1, outputTokens: 1, totalTokens: 2 },
+    finishReason: { unified: "stop", raw: "stop" },
+    usage: {
+      inputTokens: { total: 1, noCache: 1, cacheRead: 0, cacheWrite: 0 },
+      outputTokens: { total: 1, text: 1, reasoning: 0 },
+    },
     content: [{ type: "text", text: "ok" }],
     warnings: [],
   }),
diff --git a/javascript/src/voice/adapters/__tests__/elevenlabs.test.ts b/javascript/src/voice/adapters/__tests__/elevenlabs.test.ts
index e02dfd1fc..1b97da529 100644
--- a/javascript/src/voice/adapters/__tests__/elevenlabs.test.ts
+++ b/javascript/src/voice/adapters/__tests__/elevenlabs.test.ts
@@ -54,8 +54,8 @@ const isUserAudioChunk = (f: Record<string, unknown>): boolean =>
   typeof f.user_audio_chunk === "string";
 
 /**
- * Hand-rolled LanguageModelV3 stub — we only need `doGenerate`. Cast
- * through `unknown` so we don't drag the full LanguageModelV3 type
+ * Hand-rolled LanguageModelV4 stub — we only need `doGenerate`. Cast
+ * through `unknown` so we don't drag the full LanguageModelV4 type
  * surface (and its myriad provider-specific fields) into test code.
  */
 function makeFakeLanguageModel(text: string): {
@@ -64,15 +64,18 @@ function makeFakeLanguageModel(text: string): {
 } {
   const calls = { current: 0 };
   const model = {
-    specificationVersion: "v3" as const,
+    specificationVersion: "v4" as const,
     provider: "fake",
     modelId: "fake-model",
     supportedUrls: {},
     doGenerate: async () => {
       calls.current += 1;
       return {
-        finishReason: "stop",
-        usage: { inputTokens: 1, outputTokens: 1, totalTokens: 2 },
+        finishReason: { unified: "stop", raw: "stop" },
+        usage: {
+          inputTokens: { total: 1, noCache: 1, cacheRead: 0, cacheWrite: 0 },
+          outputTokens: { total: 1, text: 1, reasoning: 0 },
+        },
         content: [{ type: "text", text }],
         warnings: [],
       };
diff --git a/javascript/src/voice/adapters/composable.ts b/javascript/src/voice/adapters/composable.ts
index 5af9ba5f3..d20caf095 100644
--- a/javascript/src/voice/adapters/composable.ts
+++ b/javascript/src/voice/adapters/composable.ts
@@ -194,6 +194,9 @@ export class ComposableVoiceAgent extends VoiceAgentAdapter {
       const result = await generateText({
         model: this.llm,
         messages: this.history,
+        // history[0] is the system prompt; when the agent opens the call it is
+        // the only message, and AI SDK 7 refuses empty messages.
+        allowSystemInMessages: true,
       });
       const responseText = result.text ?? "";
       this.lastLlmResponse = responseText;
diff --git a/specs/ai-sdk-7-system-prompts.feature b/specs/ai-sdk-7-system-prompts.feature
new file mode 100644
index 000000000..3ec185070
--- /dev/null
+++ b/specs/ai-sdk-7-system-prompts.feature
@@ -0,0 +1,26 @@
+Feature: Testing agents send their system prompts the AI SDK 7 way
+  As a scenario test author on AI SDK 7
+  I want the judge and the user simulator to pass their system prompts as instructions
+  So that AI SDK 7 does not refuse their calls for carrying system messages
+
+  @unit
+  Scenario: The user simulator sends its system prompt as instructions
+    Given a user simulator agent whose model is stubbed
+    When the simulator produces the user's turn
+    Then the model call carries the system prompt as instructions
+    And no message in the model call has the system role
+
+  @unit
+  Scenario: A system turn added by the script still reaches the user simulator
+    Given a user simulator agent whose model is stubbed
+    And the transcript holds a system turn added by the script
+    When the simulator produces the user's turn
+    Then the model call allows system messages
+    And the scripted system turn is among the messages
+
+  @unit
+  Scenario: The judge sends its system prompt as instructions
+    Given a judge agent whose model is stubbed
+    When the judge decides whether to continue
+    Then the model call carries the system prompt as instructions
+    And no message in the model call has the system role