diff --git a/backend/app/services/execution_service.py b/backend/app/services/execution_service.py
index 8b70e75..2fbb29d 100644
--- a/backend/app/services/execution_service.py
+++ b/backend/app/services/execution_service.py
@@ -73,6 +73,13 @@ class ExecutionService:
         messages.append({"role": "user", "content": prompt})
 
         try:
+            # Branch by task type
+            if task.task_type == "agent_loop":
+                return ExecutionService._execute_agent_loop(
+                    db, tenant_id, task, worker, messages
+                )
+
+            # Single-call mode
             output, cost_log = LLMGateway.call(
                 worker=worker,
                 messages=messages,
@@ -141,3 +148,75 @@ class ExecutionService:
         finally:
             worker.current_task_count = max(0, worker.current_task_count - 1)
             db.commit()
+
+    @staticmethod
+    def _execute_agent_loop(db, tenant_id, task, worker, messages):
+        """Agent loop: multi-step reasoning with up to 5 iterations."""
+        max_iterations = 5
+        all_messages = list(messages)
+        final_output = ""
+        total_cost_cents = 0
+        total_tokens = 0
+        total_duration = 0
+
+        for i in range(max_iterations):
+            output, cost_log = LLMGateway.call(
+                worker=worker,
+                messages=all_messages,
+                db=db,
+                tenant_id=tenant_id,
+                project_id=task.project_id,
+                task_id=task.id,
+            )
+            total_cost_cents += cost_log.cost_cents
+            total_tokens += cost_log.total_tokens
+            total_duration += cost_log.duration_ms
+
+            # Check if agent wants to continue
+            if output.strip().startswith("NEEDS_TOOL:"):
+                tool_request = output.replace("NEEDS_TOOL:", "").strip()
+                tool_result = f"[Tool simulation] Processed: {tool_request}"
+                all_messages.append({"role": "assistant", "content": output})
+                all_messages.append({"role": "user", "content": f"Tool result: {tool_result}\nPlease continue."})
+                final_output = output
+                continue
+            else:
+                final_output = output
+                break
+
+        # Save output
+        task.output_data = json.dumps(
+            {"content": final_output, "iterations": i + 1, "total_cost": total_cost_cents},
+            ensure_ascii=False,
+        )
+
+        artifact = Artifact(
+            tenant_id=tenant_id, task_id=task.id, project_id=task.project_id,
+            title=task.title, content=final_output, artifact_type="text",
+            version=1, created_by_worker_id=worker.id,
+        )
+        db.add(artifact)
+        db.flush()
+
+        if task.requires_review:
+            review = Review(
+                tenant_id=tenant_id, task_id=task.id, project_id=task.project_id,
+                review_type="approval", status="pending",
+                review_content=final_output, submitted_at=datetime.utcnow(),
+            )
+            db.add(review)
+            task.status = "review"
+            task.review_status = "pending"
+            result_status = "needs_review"
+        else:
+            task.status = "done"
+            task.completed_at = datetime.utcnow()
+            result_status = "success"
+
+        db.commit()
+
+        return TaskExecutionResult(
+            task_id=task.id, status=result_status, output=final_output,
+            token_usage=total_tokens, cost_cents=total_cost_cents,
+            duration_ms=total_duration, artifact_id=artifact.id,
+        )