diff --git a/.changeset/eleven-bananas-grow.md b/.changeset/eleven-bananas-grow.md
new file mode 100644
index 00000000..23864f26
--- /dev/null
+++ b/.changeset/eleven-bananas-grow.md
@@ -0,0 +1,5 @@
+---
+'@openai/agents-openai': patch
+---
+
+fix: #558 prompt parameter does not work when being passed via an Agent
diff --git a/examples/basic/package.json b/examples/basic/package.json
index 5530b7f7..57fbc667 100644
--- a/examples/basic/package.json
+++ b/examples/basic/package.json
@@ -24,6 +24,7 @@
     "start:stream-text": "tsx stream-text.ts",
     "start:json-schema-output-type": "tsx json-schema-output-type.ts",
     "start:tool-use-behavior": "tsx tool-use-behavior.ts",
+    "start:prompt-id": "tsx prompt-id.ts",
     "start:tools": "tsx tools.ts",
     "start:reasoning": "tsx reasoning.ts",
     "start:local-file": "tsx local-file.ts",
diff --git a/examples/basic/prompt-id.ts b/examples/basic/prompt-id.ts
index 5bb419be..94c77281 100644
--- a/examples/basic/prompt-id.ts
+++ b/examples/basic/prompt-id.ts
@@ -4,15 +4,13 @@ async function main() {
   const agent = new Agent({
     name: 'Assistant',
     prompt: {
-      promptId: 'pmpt_684b3b772e648193b92404d7d0101d8a07f7a7903e519946',
+      promptId: 'pmpt_68d50b26524c81958c1425070180b5e10ab840669e470fc7',
       version: '1',
-      variables: {
-        poem_style: 'limerick',
-      },
+      variables: { name: 'Kaz' },
     },
   });
 
-  const result = await run(agent, 'Write about unrequited love.');
+  const result = await run(agent, 'What is your name?');
   console.log(result.finalOutput);
 }
 
diff --git a/packages/agents-openai/src/openaiResponsesModel.ts b/packages/agents-openai/src/openaiResponsesModel.ts
index bf3b0594..772a74d7 100644
--- a/packages/agents-openai/src/openaiResponsesModel.ts
+++ b/packages/agents-openai/src/openaiResponsesModel.ts
@@ -878,8 +878,8 @@ export class OpenAIResponsesModel implements Model {
     }
 
     const requestData = {
-      instructions: request.systemInstructions,
       model: this.#model,
+      instructions: normalizeInstructions(request.systemInstructions),
       input,
       include,
       tools,
@@ -1051,3 +1051,21 @@ export class OpenAIResponsesModel implements Model {
     }
   }
 }
+
+/**
+ * Sending an empty string for instructions can override the prompt parameter.
+ * Thus, this method checks if the instructions is an empty string and returns undefined if it is.
+ * @param instructions - The instructions to normalize.
+ * @returns The normalized instructions.
+ */
+function normalizeInstructions(
+  instructions: string | undefined,
+): string | undefined {
+  if (typeof instructions === 'string') {
+    if (instructions.trim() === '') {
+      return undefined;
+    }
+    return instructions;
+  }
+  return undefined;
+}
diff --git a/packages/agents-openai/test/openaiResponsesModel.test.ts b/packages/agents-openai/test/openaiResponsesModel.test.ts
index f45bdbd4..5d33d76b 100644
--- a/packages/agents-openai/test/openaiResponsesModel.test.ts
+++ b/packages/agents-openai/test/openaiResponsesModel.test.ts
@@ -13,7 +13,7 @@ describe('OpenAIResponsesModel', () => {
     setTracingDisabled(true);
   });
   it('getResponse returns correct ModelResponse and calls client with right parameters', async () => {
-    withTrace('test', async () => {
+    await withTrace('test', async () => {
       const fakeResponse = {
         id: 'res1',
         usage: {
@@ -74,8 +74,67 @@ describe('OpenAIResponsesModel', () => {
     });
   });
 
+  it('normalizes systemInstructions so empty strings are omitted', async () => {
+    await withTrace('test', async () => {
+      const fakeResponse = {
+        id: 'res-empty-instructions',
+        usage: {
+          input_tokens: 0,
+          output_tokens: 0,
+          total_tokens: 0,
+        },
+        output: [],
+      };
+      for (const systemInstructions of ['', '   ']) {
+        const request = {
+          systemInstructions,
+          input: 'hello',
+          modelSettings: {},
+          tools: [],
+          outputType: 'text',
+          handoffs: [],
+          tracing: false,
+          signal: undefined,
+        };
+        const createMock = vi.fn().mockResolvedValue(fakeResponse);
+        await new OpenAIResponsesModel(
+          { responses: { create: createMock } } as unknown as OpenAI,
+          'gpt-test',
+        ).getResponse(request as any);
+
+        expect(createMock).toHaveBeenCalledTimes(1);
+        const [args] = createMock.mock.calls[0];
+        expect('instructions' in args).toBe(true);
+        expect(args.instructions).toBeUndefined();
+      }
+
+      for (const systemInstructions of [' a ', 'foo']) {
+        const request = {
+          systemInstructions,
+          input: 'hello',
+          modelSettings: {},
+          tools: [],
+          outputType: 'text',
+          handoffs: [],
+          tracing: false,
+          signal: undefined,
+        };
+        const createMock = vi.fn().mockResolvedValue(fakeResponse);
+        await new OpenAIResponsesModel(
+          { responses: { create: createMock } } as unknown as OpenAI,
+          'gpt-test',
+        ).getResponse(request as any);
+
+        expect(createMock).toHaveBeenCalledTimes(1);
+        const [args] = createMock.mock.calls[0];
+        expect('instructions' in args).toBe(true);
+        expect(args.instructions).toBe(systemInstructions);
+      }
+    });
+  });
+
   it('merges top-level reasoning and text settings into provider data for Responses API', async () => {
-    withTrace('test', async () => {
+    await withTrace('test', async () => {
       const fakeResponse = {
         id: 'res-settings',
         usage: {
@@ -134,7 +193,7 @@ describe('OpenAIResponsesModel', () => {
   });
 
   it('getStreamedResponse yields events and calls client with stream flag', async () => {
-    withTrace('test', async () => {
+    await withTrace('test', async () => {
       const fakeResponse = { id: 'res2', usage: {}, output: [] };
       const events: ResponseStreamEvent[] = [
         { type: 'response.created', response: fakeResponse as any },