diff --git a/packages/agent/src/agent-loop.ts b/packages/agent/src/agent-loop.ts
index 9e612d05d..795a8d59c 100644
--- a/packages/agent/src/agent-loop.ts
+++ b/packages/agent/src/agent-loop.ts
@@ -205,7 +205,13 @@ async function runLoop(
 			const toolResults: ToolResultMessage[] = [];
 			hasMoreToolCalls = false;
 			if (toolCalls.length > 0) {
-				const executedToolBatch = await executeToolCalls(currentContext, message, config, signal, emit);
+				// A "length" stop means the output was cut off by the token limit, so
+				// every tool call in the message may carry truncated arguments. Fail
+				// them all instead of executing potentially borked calls.
+				const executedToolBatch =
+					message.stopReason === "length"
+						? await failToolCallsFromTruncatedMessage(toolCalls, emit)
+						: await executeToolCalls(currentContext, message, config, signal, emit);
 				toolResults.push(...executedToolBatch.messages);
 				hasMoreToolCalls = !executedToolBatch.terminate;
 
@@ -367,6 +373,40 @@ async function streamAssistantResponse(
 	return finalMessage;
 }
 
+/**
+ * Fail all tool calls from an assistant message that was truncated by the
+ * output token limit. Streamed tool-call arguments are finalized with a
+ * best-effort JSON salvage parser, so a truncated message can yield tool calls
+ * whose arguments parse and validate but are silently incomplete. None of them
+ * are safe to execute; report each as an error so the model can re-issue them.
+ */
+async function failToolCallsFromTruncatedMessage(
+	toolCalls: AgentToolCall[],
+	emit: AgentEventSink,
+): Promise<ExecutedToolCallBatch> {
+	const messages: ToolResultMessage[] = [];
+	for (const toolCall of toolCalls) {
+		await emit({
+			type: "tool_execution_start",
+			toolCallId: toolCall.id,
+			toolName: toolCall.name,
+			args: toolCall.arguments,
+		});
+		const finalized: FinalizedToolCallOutcome = {
+			toolCall,
+			result: createErrorToolResult(
+				`Tool call "${toolCall.name}" was not executed: the response hit the output token limit, so its arguments may be truncated. Re-issue the tool call with complete arguments.`,
+			),
+			isError: true,
+		};
+		await emitToolExecutionEnd(finalized, emit);
+		const toolResultMessage = createToolResultMessage(finalized);
+		await emitToolResultMessage(toolResultMessage, emit);
+		messages.push(toolResultMessage);
+	}
+	return { messages, terminate: false };
+}
+
 /**
  * Execute tool calls from an assistant message.
  */
diff --git a/packages/agent/test/agent-loop.test.ts b/packages/agent/test/agent-loop.test.ts
index 62eadf810..038680478 100644
--- a/packages/agent/test/agent-loop.test.ts
+++ b/packages/agent/test/agent-loop.test.ts
@@ -307,6 +307,79 @@ describe("agentLoop with AgentMessage", () => {
 		}
 	});
 
+	it("should not execute tool calls from a length-truncated assistant message", async () => {
+		const toolSchema = Type.Object({ value: Type.String() });
+		const executed: string[] = [];
+		const tool: AgentTool<typeof toolSchema, { value: string }> = {
+			name: "echo",
+			label: "Echo",
+			description: "Echo tool",
+			parameters: toolSchema,
+			async execute(_toolCallId, params) {
+				executed.push(params.value);
+				return {
+					content: [{ type: "text", text: `echoed: ${params.value}` }],
+					details: { value: params.value },
+				};
+			},
+		};
+
+		const context: AgentContext = {
+			systemPrompt: "",
+			messages: [],
+			tools: [tool],
+		};
+
+		const config: AgentLoopConfig = {
+			model: createModel(),
+			convertToLlm: identityConverter,
+		};
+
+		let callIndex = 0;
+		const streamFn = () => {
+			const stream = new MockAssistantStream();
+			queueMicrotask(() => {
+				if (callIndex === 0) {
+					// Output hit the token limit mid tool call. The salvage parser can
+					// produce arguments that validate but are silently truncated, so
+					// nothing in this message may execute.
+					const message = createAssistantMessage(
+						[{ type: "toolCall", id: "tool-1", name: "echo", arguments: { value: "hel" } }],
+						"length",
+					);
+					stream.push({ type: "done", reason: "length", message });
+				} else {
+					const message = createAssistantMessage([{ type: "text", text: "done" }]);
+					stream.push({ type: "done", reason: "stop", message });
+				}
+				callIndex++;
+			});
+			return stream;
+		};
+
+		const events: AgentEvent[] = [];
+		const stream = agentLoop([createUserMessage("echo something")], context, config, undefined, streamFn);
+		for await (const event of stream) {
+			events.push(event);
+		}
+
+		// The tool must never execute with potentially truncated arguments.
+		expect(executed).toEqual([]);
+
+		const toolEnd = events.find((e) => e.type === "tool_execution_end");
+		expect(toolEnd).toBeDefined();
+		if (toolEnd?.type === "tool_execution_end") {
+			expect(toolEnd.isError).toBe(true);
+			const text = toolEnd.result.content.find((c: { type: string }) => c.type === "text");
+			expect(text && "text" in text ? text.text : "").toContain("output token limit");
+		}
+
+		// The loop continues so the model can re-issue the tool call.
+		expect(callIndex).toBe(2);
+		const messages = await stream.result();
+		expect(messages[messages.length - 1].role).toBe("assistant");
+	});
+
 	it("should execute mutated beforeToolCall args without revalidation", async () => {
 		const toolSchema = Type.Object({ value: Type.String() });
 		const executed: Array<string | number> = [];