diff --git a/.changeset/recover-invalid-tool-input.md b/.changeset/recover-invalid-tool-input.md
new file mode 100644
index 0000000000..65be97da97
--- /dev/null
+++ b/.changeset/recover-invalid-tool-input.md
@@ -0,0 +1,7 @@
+---
+'@workflow/ai': patch
+---
+
+DurableAgent: recover from invalid tool-call input instead of aborting the stream
+
+When a model emits a tool call whose arguments fail `inputSchema` validation (and no `experimental_repairToolCall` fixes it), `executeTool` now returns the validation error to the model as an `error-text` tool result — the same way tool *execution* errors are already handled — instead of throwing and aborting the whole agent stream. In a durable workflow that throw fails the entire run, so a single occasionally-malformed model tool-call could kill a long-running task with no chance for the agent to self-correct. The agent now sees the error as a tool result and can fix the arguments and retry within its step budget.
diff --git a/packages/ai/src/agent/durable-agent.test.ts b/packages/ai/src/agent/durable-agent.test.ts
index f78346ade2..2f3ff01343 100644
--- a/packages/ai/src/agent/durable-agent.test.ts
+++ b/packages/ai/src/agent/durable-agent.test.ts
@@ -2006,6 +2006,82 @@ describe('DurableAgent', () => {
       });
     });
 
+    it('should convert invalid tool input to error-text result instead of failing stream', async () => {
+      const tools: ToolSet = {
+        strictTool: {
+          description: 'A tool with a strict input schema',
+          inputSchema: z.object({ requiredField: z.string().min(1) }),
+          execute: async () => ({ ok: true }),
+        },
+      };
+
+      const mockModel = createMockModel();
+
+      const agent = new DurableAgent({
+        model: async () => mockModel,
+        tools,
+      });
+
+      const mockWritable = new WritableStream({
+        write: vi.fn(),
+        close: vi.fn(),
+      });
+
+      const mockMessages: LanguageModelV3Prompt = [
+        { role: 'user', content: [{ type: 'text', text: 'test' }] },
+      ];
+
+      const { streamTextIterator } = await import('./stream-text-iterator.js');
+      const mockIterator = {
+        next: vi
+          .fn()
+          .mockResolvedValueOnce({
+            done: false,
+            value: {
+              toolCalls: [
+                {
+                  toolCallId: 'test-call-id',
+                  toolName: 'strictTool',
+                  // Valid JSON, but violates the schema (empty string fails .min(1)).
+                  input: '{"requiredField":""}',
+                } as LanguageModelV3ToolCall,
+              ],
+              messages: mockMessages,
+            },
+          })
+          .mockResolvedValueOnce({ done: true, value: [] }),
+      };
+      vi.mocked(streamTextIterator).mockReturnValue(
+        mockIterator as unknown as MockIterator
+      );
+
+      // Invalid tool input should be handled gracefully, not reject the stream.
+      await expect(
+        agent.stream({
+          messages: [{ role: 'user', content: 'test' }],
+          writable: mockWritable,
+        })
+      ).resolves.not.toThrow();
+
+      // Verify the validation error was sent back as an error-text tool result
+      // (so the model can correct its arguments and retry).
+      expect(mockIterator.next).toHaveBeenCalledTimes(2);
+      const toolResultsCall = mockIterator.next.mock.calls[1][0];
+      expect(toolResultsCall).toBeDefined();
+      expect(toolResultsCall).toHaveLength(1);
+      expect(toolResultsCall[0]).toMatchObject({
+        type: 'tool-result',
+        toolCallId: 'test-call-id',
+        toolName: 'strictTool',
+        output: {
+          type: 'error-text',
+        },
+      });
+      expect(toolResultsCall[0].output.value).toContain(
+        'Invalid input for tool "strictTool"'
+      );
+    });
+
     it('should call onFinish with steps and messages when streaming completes', async () => {
       const mockModel = createMockModel();
 
diff --git a/packages/ai/src/agent/durable-agent.ts b/packages/ai/src/agent/durable-agent.ts
index b3fbfe6e3c..c96ba9183a 100644
--- a/packages/ai/src/agent/durable-agent.ts
+++ b/packages/ai/src/agent/durable-agent.ts
@@ -1656,7 +1656,21 @@ async function executeTool(
         );
       }
     }
-    throw parseError;
+    // Input that fails to parse or validate (even after repair) is recoverable,
+    // exactly like a tool execution error below: feed the error back to the model
+    // as an error-text result so the agent can correct the call and retry, instead
+    // of aborting the entire stream. This aligns with AI SDK's streamText behavior
+    // for tool failures. Reaches here both for malformed JSON and for the
+    // re-thrown "Invalid input for tool ..." schema-validation error above.
+    return {
+      type: 'tool-result' as const,
+      toolCallId: toolCall.toolCallId,
+      toolName: toolCall.toolName,
+      output: {
+        type: 'error-text' as const,
+        value: getErrorMessage(parseError),
+      },
+    };
   }
 
   return recordSpan({