《吻我之前，先把灯关了》 "Turn Off the Lights Before You Kiss Me" 据专家研究分析，恐怖片和爱情动作片，好莱坞惊险片都能刺激大脑分泌多巴胺，我现在把它们都放在一起了，那是不是应该双倍快乐来了。 According to expert research and analysis, horror movies, romantic action films, and Hollywood thrillers can all stimulate the brain to release dopamine. Since I’ve combined them all together, does that mean I’m in for a double dose of happiness? grok imagine prompt: "Turn Off the Lights Before You Kiss Me" 15 seconds, 9:16 vertical format, first-person perspective, recorded on a smartphone, single continuous take. [Realistic Cinematography] The footage is captured entirely by the rear-facing camera of the smartphone held by the male protagonist; no other cameras are involved. The male protagonist holds the phone in his right hand, using the primary (1×) rear lens—providing a field of view equivalent to approximately 26mm—at 1080p resolution and 30fps. The device is positioned between his chest and chin (at a height of roughly 1.45 meters), aimed toward a woman standing in the bathroom doorway. His left hand occasionally enters the frame to operate the light switch. The smartphone itself—its body, back panel, camera module, screen, or flash—is never visible in the shot. Only the light cast by the phone onto the subject, her clothing, and the walls can be seen. There are no mirrors present; this is not a mirror-based selfie. The footage retains the raw, authentic texture of smartphone video—no beauty filters, color grading, or cinematic effects. Skin details such as pores, fine beads of sweat, stray hairs, and natural skin-tone variations are visible. There is slight digital noise in the shadows, and brief motion blur occurs during rapid hand movements or moments of physical instability. Subtle camera shake stems from the holder’s breathing and wrist movements. Changes in proximity are achieved through actual physical steps or the subject pulling the holder closer; there is no digital zooming, automated panning, reverse-angle cutting, or slow motion. The focus remains primarily on the woman’s face, with only momentary hesitation when the distance shifts abruptly—there is no persistent out-of-focus effect. [Characters and Attire] The woman in the doorway (@80130fc6-f744-4f61-9faf-f630a1eb8638) is 28 years old. Her appearance—including her black hair styled in an updo, purple floral hair accessory, earrings, long-sleeved Japanese-style (wafuku) garment in purple (@d7529a1b-8cdb-41e5-816e-c59aa4051803), and dark-colored *obi* (sash)—is consistent with the video uploaded by the user. Her clothing remains intact; the neckline stays in place, and she does not scratch her own chest or tug at her garments. Leaning slightly against the doorframe, she keeps her shoulders relaxed and lowered, her neck elongated, and her legs positioned one in front of the other—creating a natural, subtle S-curve in her posture. She gazes toward the male protagonist with half-lidded eyes and a faint, inviting smile, as if beckoning him closer. From the very beginning, her right hand holds a small pinch of fabric from the male protagonist’s left cuff. This point of contact—her right hand gripping his left cuff—is situated in the lower-left portion of the frame and must remain visible. Because she is only holding the fabric rather than clamping down on his wrist, he retains the freedom to lift his left hand and operate the light switch. Initially, her left hand is concealed between her body and the doorframe. The grayish-white fingers that appear later belong to this same hand; the creature possesses only two hands throughout the entire sequence. The real girlfriend is located in the hallway behind the male protagonist. She is present only as a voice calling out from off-screen; she is never shown on camera, and no second female face appears. 【Environment and Space】 The scene takes place at the entrance to a bathroom in an old apartment building late at night. The setting—featuring aged off-white tiles, a worn wooden doorframe, a standard wall-mounted light switch, and a towel hanging inside the bathroom—conveys the lived-in, weathered feel of an old residence, rather than a ruin. The woman is positioned slightly to the right of the frame’s center, with the open bathroom door behind her. The light switch is located on the wall to the left, approximately thirty centimeters from the male protagonist’s left hand. It is visible from the start and does not suddenly appear later. Behind the male protagonist lies a hallway. The bathroom door remains open inward; the camera angle stays consistent throughout, the door does not move on its own, and the room’s layout remains unchanged. At the start, the woman’s face is about sixty centimeters from the camera. After she gives a gentle tug on his cuff, the male protagonist shifts forward slightly, reducing the distance to approximately forty-five centimeters. The distance closes further only during the final dragging motion. 【Ambient Lighting】 At the beginning, a warm-white overhead light in the bathroom is on, illuminating the woman’s head, shoulders, and the doorway, while the hallway remains dim. The LED flash on the back of the smartphone is illuminated from the very first frame; it remains on and in the same position throughout the entire 15-second duration. Up close, the face is illuminated by the harsh, direct light of a standard mobile phone; small shadows form beside the nose and under the chin, and there are localized highlights on the fabric of the clothing—no artificial "beauty" fill lighting is applied. The only light the male lead turns off is the overhead bathroom light. Once the overhead light is extinguished, the bathroom background remains dim and does not brighten on its own. The phone’s light continues to illuminate her face and the hand gripping the sleeve cuff. The camera’s auto-exposure adjusts only slightly, failing to brighten the dark bathroom to a daylight-like state. When the "mask" (the outer layer of the face) is peeled back, it is illuminated by the same mobile phone light as the rest of the face. The space deep behind the shell remains dark, and the exposed eyeball shows only natural, subtle reflections—it does not emit its own light. [The "False Face" Effect] The horror stems from a thin, pliable, artificial layer being peeled back from within—it is not a case of natural skin being torn or a mouth stretching into an impossible, elongated shape. A grayish-white hand—with only two or three fingertips visible—emerges from the corner of her mouth, hooks onto the lower edge of the "mask," and peels half of the face upward and to the side. The "mask" has a distinct thickness and flexibility; the cheek and part of the area beside the nose fold back, rather than stretching into long, tubular, funnel-like, or tongue-like shapes. Behind the peeled-back section lies a deep, dark cavity containing a normal-sized human eye, which turns to look directly at the camera. The other half of the woman’s face remains intact, establishing the premise that there is an artificial layer on the outside, with something else concealed beneath it. There is no blood, gore, exposed muscle, black ooze, or liquid-like melting effect. The grayish hand does not reach in from off-screen; its fingertips must first emerge from within the corner of the mouth. [Strict 15-Second Timeline] 0–4 Seconds | Gripping the sleeve, an invitation to draw closer. The woman leans sideways against the doorframe, looking first into the male lead’s eyes, then glancing briefly toward the light switch on the left. With her right hand, she grasps the male lead’s left sleeve cuff—positioned at the bottom-left of the frame—and gives it a gentle tug toward herself. The fabric of the cuff pulls taut, prompting the male lead to take a small step forward; the camera moves in tandem with his approach. She lifts her chin slightly and says in a low voice: “The lights are too bright.” Her tone is intimate, with a suppressed smile playing at the corners of her mouth. There is no sign of a monstrous visage, nor does she glance toward her own mouth beforehand. Her right hand continues to hold the same spot on his sleeve; she does not shift her grip to his lapel. 4–7 seconds | Flipping the switch; the background darkens The male protagonist reaches out with his left hand toward a light switch already visible on the left side of the frame. She keeps her grip on his left sleeve, causing her right arm to extend naturally as he moves—she neither lets go nor suffers any "clipping" (graphical glitching). His fingertips press the switch, producing a distinct *click*. The bathroom overhead light goes out, and the tiles, towels, and the space behind the woman grow noticeably darker. Her face remains illuminated by the fill light from a mobile phone. The male protagonist’s left hand pulls away from the switch and returns to the space between them. The camera tilts slightly with the movement, then steadies. 7–9 seconds | A warning from behind From about two meters behind the camera, the man’s actual girlfriend calls out urgently, in an identical tone: “Don’t turn off the light!” The woman in front of him says nothing; her lip movements do not correspond to this warning. As the male protagonist begins to turn, the camera shifts slightly to the side—but the hand gripping his sleeve suddenly tightens, holding him in place. Her smile vanishes. Her eyes glance past the camera toward the space behind the man, then return to the lens. Her knuckles whiten as she grips the sleeve, causing deeper creases to form in the fabric. 9–12 seconds | Half of her face is peeled back from within Two or three grayish-white fingertips emerge from the inner corner of her mouth. The fingertips hook onto the lower edge of the "shell" of her face, slowly peeling it upward and outward. The corner of her mouth and her cheek—forming a single, thin, pliable layer—fold back together; the lips do not stretch or elongate independently. Half of the artificial face is lifted away, revealing an eye in the darkness beneath that shifts to look directly at the camera. The intact half of her face remains visible alongside the eye revealed from within. For a brief moment, the camera steadies, allowing at least half a second to take in the details: Fingers emerge from within the face; the "shell" of the face begins to fold back, revealing another eye underneath. The woman’s right hand maintains a firm grip on the man’s left cuff—she cannot let go, even as she peels back the face. The only sounds are the dry, grating friction of the mask and the man’s sudden, sharp intake of breath. 12–15 seconds | The grip shifts into a forceful tug. The man immediately grabs the woman’s right wrist—the one holding his cuff—and tries to pry it away. The point of contact is clearly visible; he does not reach for the grey, peeling fingers. The creature’s right arm suddenly retracts toward the bathroom, yanking the cuff hard. The sequence must be clear: The cuff goes taut → the man’s left arm is pulled straight → his body stumbles forward → the camera (held in his right hand) is dragged closer. The camera moves from about 45 centimeters away to roughly 15 centimeters, accompanied by a sudden dip and tilt—this is not a smooth, controlled zoom. The partially peeled-back face, the unblinking eye beneath, and the greyish-white fingers now dominate the frame. The other half of the face remains "normal," preserving its original appearance. In the final half-second, the man collides with the wooden doorframe, producing a dull thud. The camera jerks to the side; the edge of the facial shell sweeps across the lens, partially obscuring the view. The recording ends abruptly while the struggle is still underway. No fade to black, no additional dialogue, and no new creatures appear. [Ambient Audio] Captured via a mobile phone microphone: close-range speech, the friction of fabric at the cuff, the sound of a light switch, the hum of a bathroom exhaust fan, the distant ambient noise of the city, heavy breathing, and the final impact against the doorframe. The exhaust fan continues to run independently; it does not stop when the overhead light is switched off. The background sounds maintain a sense of spatial distance—they are not suddenly muted for the sake of a "jump scare." The opening line, "The light is too bright," comes from the woman in front of him. The mid-sequence line, "Don't turn off the light," clearly originates from behind the man; the woman in front does not move her lips. There is no further dialogue—no sinister laughter, voice-over narration, heartbeat sound effects, or pre-recorded horror audio cues. [Mandatory Continuity] The only camera is the mobile phone through which the audience is viewing the scene; the phone itself is never visible within the frame, nor is there any additional camera operator. The male protagonist holds the phone in his right hand; he uses his left hand to flip the light switch, then immediately grabs the monster’s right wrist. The monster’s right hand remains firmly gripping the same spot on the male protagonist’s left cuff; its left hand lifts a portion of the "false face" from underneath. Do not generate a third hand, swap left and right hands, or have the monster’s hand shift its grip from the cuff to the female protagonist’s own clothing. Once the bathroom light is turned off, it remains off; the mobile phone’s light source (fill light) stays on. The positions of the door and the light switch remain fixed. Changes in distance must be achieved through physical movement (stepping or dragging); do not use digital zoom. The "false face" should be lifted, not stretched like rubber; do not generate long tongues, tubular mouths, an elongated jaw, or a melting face.

AWS 资深专家解决方案架构师 Daniel Abib 在官方博客里算了一笔账：同步调用时，Lambda 的计费时长几乎等于 agent 思考的时长，等待本身在按全量计算计费。 《在无服务器流水线中异步调用 Amazon Bedrock AgentCore agent 的模式》 在无服务器流水线里异步调用 Amazon Bedrock AgentCore agent，可以消除你的 AI agent 处理请求期间的空转计算成本。一个常见的例子是文档校验：在房地产融资的后台办公室里，agent 可以读取房产记录或贷款合同，推理信息是否完整、一致，然后返回一个下游步骤据此行动的结论。Amazon Bedrock AgentCore 提供了一个平台，让你能用任意框架或模型大规模地构建、连接和优化 agent。 这类 agent 带来一个传统流水线步骤没有的特性：它们在回答之前要思考一会儿。思考多久取决于提示词、模型和文档，但很少是即时的，而这个延迟改变了你调用它的方式。最常见的首版实现是用一个计算服务（比如 AWS Lambda 函数）调用 agent 并等待响应。函数等待期间什么都不做，但它仍在运行，每一秒都在计费。 值得看清成本到底落在哪里，因为调用的两侧计费方式不同。Amazon Bedrock AgentCore runtime（Amazon Bedrock AgentCore 的一项能力）采用基于消耗的计费模型，agent 空闲时不收 CPU 费用。例如，在等待大语言模型生成响应、或等待工具或 Model Context Protocol（MCP）调用返回的这段时间里，你只按内存计费，不按 CPU 计费。调用 agent 的那个计算服务没有这种特性。发出同步调用的 Lambda 函数、容器或 Amazon EC2 实例会一直阻塞，在 agent 响应之前始终持有（并支付）完整的计算配额。所以浪费不在 agent 一侧，而在调用方——它挂在一个打开的连接上空转。 这就让调用方的成本跟着 agent 的运行时间走。阻塞在 agent 上的函数，计费时长几乎等于整个处理时间；而启动 agent 后立即返回的函数，只按短暂的派发计费。解法是在等待期间释放调用方的计算资源，等 agent 出了结果再恢复流水线。本文展示三种做到这一点的模式（task-token 回调、直接服务集成、durable function），并与阻塞式反模式对比。 示例流水线 为了让各模式在同等条件下比较，我们把每个模式都跑在同一条流水线上，只更换调用 agent 的那一步。这条流水线是一个刻意简单、虚构的场景（房地产融资文档校验），选它只是为了把编排讲清楚。它不是本文的重点，它代表任何先调用 agent（或其他慢服务）、再对结果采取行动的流程，你可以把自家用例代入。 流水线有五个阶段： 1. Extract：一个 Lambda 函数对文档做 OCR 和文本提取。（提取是模拟的，所以场景不需要真实文档。） 2. Identify：一个 Lambda 函数对文档分类并设置路由标志（shouldOrganize、shouldValidate）。 3. Route：一个 Choice 状态根据这些标志决定流程走向。 4. Organize and Validate：一个 Parallel 状态在组织文档的同时，让 Amazon Bedrock AgentCore agent 在另一个分支里做校验。这个 Validate 分支是各模式之间唯一变化的部分。 5. Result：一个 Lambda 函数处理 agent 的结论，决定下一步动作（批准，或退回修改）。 图示如下：水平流程依次是 Start、Extract、Identify、Route（Choice）、并行的 Organize-and-Validate 阶段、Result、End，Validate 分支被标为可更换，图例列出了它的四种实现方式：阻塞反模式、模式 1（task token）、模式 2（直接集成）、模式 3（durable function）。流水线本身在每种情况下保持不变，只有 Validate 分支被换掉来演示各自的调用模式。 同一个 Amazon Bedrock AgentCore agent 服务全部四种情况。agent 检查每次调用并选择如何响应：如果收到 AWS Step Functions 的 task token，就在完成时唤醒那个执行；如果收到 durable function 的 callback ID，就唤醒那个 durable function；两者都没收到，就直接在响应里返回结论。这意味着你可以更换编排模式，而不必修改或重新部署 agent。 agent 如何不阻塞调用方就把控制权交回来 机制是 agent 的 action group 里的一个"归还控制权"动作。agent 推理结束后，调用一个 Lambda，把结果和 task token 发回 Step Functions。（后面的模式 2 会让 Step Functions 直接与 AgentCore 集成，完全省掉这个 Lambda。） 下面这段代码是这个 Lambda 的核心： The tool the agent calls once it reaches a verdict @tool def conclude_validation(approved: bool, issues: list, summary: str) -> str: verdict = {"approved": approved, "issues": issues, "summary": summary, "source": "agentcore"} # A Step Functions task token was passed: resume that execution if task_token: sfn.send_task_success(taskToken=task_token, output=json.dumps(verdict)) return "Step Functions resumed." # A durable-function callback ID was passed: resume the durable function if callback_id: lambda_client.send_durable_execution_callback_success( CallbackId=callback_id, Result=json.dumps(verdict).encode("utf-8")) return "Durable function resumed." # Neither was passed: this is a synchronous call, return the verdict inline return "Verdict recorded." 入口函数根据同样的信号决定在后台跑还是同步跑： @app.async_task async def validate_document_async(prompt, document, extracted_text): # Background work; conclude_validation fires the right callback when done agent = build_agent() await agent.invoke_async(message(prompt, document, extracted_text)) @app.entrypoint async def handler(event): task_token = event.get("taskToken") # passed by the task-token pattern callback_id = event.get("callbackId") # passed by the durable-function pattern # Asynchronous: start the work and return "accepted" right away if task_token or callback_id: asyncio.create_task(validate_document_async(...)) return {"status": "accepted"} # Synchronous: run now and return the verdict in the response agent = build_agent() await agent.invoke_async(message(...)) return verdict agent 就位之后，本文剩下的部分聚焦于调用它的四种方式。代码和基础设施定义都是示例中的摘录，用来演示每种模式。 调用 agent：四种方式 先讲阻塞式反模式以确立基准成本，再展示避开它的三个模式。 阻塞式反模式 最直接的实现：在同一个 Lambda 函数里调用 agent 并等待答案。它能用，而且实现简单——这正是它如此常见的原因——但函数在 agent 思考的整个期间一直活着。 // The Lambda function blocks here until the agent responds const response = await agentcore.send( new InvokeAgentRuntimeCommand({ agentRuntimeArn: AGENT_RUNTIME_ARN, payload: new TextEncoder().encode(JSON.stringify(payload)), runtimeSessionId: sessionId, }) ); // The function stays alive and billed for the entire time the agent is thinking. 函数的计费时长最终大约等于 agent 的处理时间。后面三个模式消除这段空闲成本，各自做出不同的取舍。特别是模式 2 使用 Step Functions 对 AgentCore Harness (InvokeHarness) 的优化集成，完全去掉 Lambda。 模式 1：task-token 回调 + 派发函数 这个模式在路径里保留一个 Lambda 做自定义逻辑，但去掉空闲成本。Step Functions 用 waitForTaskToken 集成调用这个函数——传入一个 task token 后暂停执行。函数用这个 token 启动 agent，然后在几秒内返回。执行保持暂停，不产生任何计算计费，直到 agent 用这个 token 调用 SendTaskSuccess 恢复它。 // Start the agent, pass the task token, and return without waiting const response = await agentcore.send( new InvokeAgentRuntimeCommand({ agentRuntimeArn: AGENT_RUNTIME_ARN, payload: new TextEncoder().encode(JSON.stringify({ ...payload, taskToken })), runtimeSessionId: sessionId, }) ); // Returning here does not complete the step. Step Functions stays paused until // the agent calls SendTaskSuccess with this task token. return { dispatched: true }; 对应的状态从上下文取出 token，并设置超时和心跳作为安全网——这样沉默的 agent 会让执行干净地失败，而不是无限期停在那里： "ValidateDispatch": { "Type": "Task", "Resource": "arn:aws:states:::lambda:invoke.waitForTaskToken", "Parameters": { "FunctionName": "${ValidateDispatcherFunctionArn}", "Payload": { "taskToken.$": "$$.Task.Token", "document.$": "$.document", "extractedText.$": "$.extract.extractedText", "executionId.$": "$$.Execution.Id" } }, "TimeoutSeconds": 120, "HeartbeatSeconds": 60, "Next": "AgentCoreValidation" } 成本。 Lambda 函数仍然会跑，但只跑到启动 agent 并返回为止：几秒钟，无论 agent 之后要跑多久。你只为这次短暂派发付费，不为等待付费——因为函数在 agent 干活时已经关掉了。等待由暂停的 Step Functions 执行承担，它不为空闲计算计费。这是与阻塞版本的关键区别：阻塞版本里函数的计费时间跟着 agent 的处理时间走。 模式 2：直接服务集成 当你不需要围绕 agent 调用写自定义代码时，可以去掉派发函数，让 Lambda 完全离开路径。Step Functions 可以通过 AWS SDK 服务集成直接调用 Amazon Bedrock AgentCore，所以 agent 的响应直接流入下一个状态。Validate 分支于是变成单个 Task 状态： "ValidateDirect": { "Type": "Task", "Resource": "arn:aws:states:::aws-sdk:bedrockagentcore:invokeAgentRuntime", "Parameters": { "AgentRuntimeArn": "${AgentRuntimeArn}", "RuntimeSessionId.$": "States.Hash($$.Execution.Id, 'SHA-256')", "Payload.$": "States.JsonToString($.prep.agentInput)" }, "ResultSelector": { "raw.$": "$.Response" }, "TimeoutSeconds": 120, "Next": "ParseVerdict" } 成本。 路径里没有 Lambda，所以没有空闲的 Lambda 计算可付。Step Functions 承担等待，Standard 工作流按状态转换计费而不是按等待时长计费——处理期间的实质成本只有 agent 本身。 模式 3：Lambda durable function 如果你更愿意用代码而不是状态机来表达编排，Lambda durable function 给你同样的成本行为。用 @aws/durable-execution-sdk-js SDK，流水线阶段变成 context.step 调用，并行工作变成 context.parallel，等待 agent 变成 context.waitForCallback。等待期间函数挂起，不计计算费。agent 用 SendDurableExecutionCallbackSuccess 恢复它。 // Suspend the function until the agent calls back const result = await ctx.waitForCallback( "validate-agentcore", async (callbackId) => dispatchAgentCore(callbackId, document, extractedText, executionId), { timeout: { seconds: 120 } } ); 成本。 一个函数承载整条流水线，但挂起等待 agent 期间不按计算计费。你只为挂起之间短暂爆发的那几次执行付费——与 task-token 模式相同的经济性——而不是为等待付费。 测量差异 重点不是某个具体数字。agent 的运行时间随提示词、模型和文档变化。重要的是 task-token 模式里两个值之间的关系：Validate 状态活跃了多久，对比派发函数实际被计费了多久。我们测试中的单次运行让这个关系可见： Step Functions, ValidateDispatch state Returned (TaskSubmitted): 14:08:19 <- the function returned and shut down Resumed (TaskSucceeded): 14:08:34 <- the agent woke the execution State active ........ 19.6s Lambda, dispatcher function (CloudWatch REPORT) Billed duration ..... 4.8s Result: the state was active for 19.6s, but the function was billed for 4.8s. The ~14.8s in between is wait time with no Lambda function running. 在 Step Functions 的事件历史里也能看到同样的事：task-token 模式下，TaskSubmitted 事件（函数返回）与 TaskSucceeded（agent 恢复流程）之间隔着 agent 的处理时间。同步调用没有这样的间隔。 下表总结了每种模式下 Validate 分支（前面图示中高亮的部分）的实现方式： • 编排器：阻塞＝Step Functions；模式 1＝Step Functions；模式 2＝Step Functions；模式 3＝Lambda（代码） • 路径中的 Lambda：阻塞＝有，存活且计费；模式 1＝有，但提前返回；模式 2＝无；模式 3＝durable function 本身（挂起） • 等待期间的空闲 Lambda 计算：阻塞＝为整个等待付费；模式 1＝无；模式 2＝无；模式 3＝无 • agent 调用周围的自定义代码：阻塞＝有；模式 1＝有；模式 2＝有限（状态转换）；模式 3＝有 • 调用方与 agent 解耦：阻塞＝否；模式 1＝是；模式 2＝否；模式 3＝是 • 相对复杂度：阻塞＝最低；模式 1＝较高（token 和回调的 IAM）；模式 2＝最低；模式 3＝中等（检查点/重放） 读这张表时有一个提醒：总流水线时长不是有用的比较指标，因为它被 agent 自己的推理时间主导——每次运行都不同，而且在所有场景里基本相同。有意义的差异是等待期间你为多少计算付了钱，这由上表和计费时长读数体现。agent 运行时间增长时，派发函数的计费时间保持平稳。上面的数字来自单次运行，把它们当作这个关系的示意而非基准，请测量你自己的负载。 如何选择模式 下表总结了取舍，帮你选择模式： • 调用方成本：阻塞＝整个 agent 处理时间；模式 1＝几秒（仅派发）；模式 2＝零（无 Lambda）；模式 3＝几秒（仅派发） • 集成工作量：阻塞＝低；模式 1＝中高（IAM、心跳、超时）；模式 2＝低（单个 Task 状态）；模式 3＝中等（检查点-重放模型） • 业务逻辑位置：阻塞＝在 Lambda（前 + 后）；模式 1＝在 Lambda（前 + 后）；模式 2＝只在 Amazon States Language（ASL）（内置函数）；模式 3＝在 Lambda（顺序代码） • 最适合：阻塞＝原型、短 agent；模式 1＝自定义前后处理逻辑；模式 2＝纯编排、无自定义代码；模式 3＝单个函数里的复杂异步工作流 最佳实践 防 agent 永不回答。 给每个 waitForTaskToken 状态设置 TimeoutSeconds，让执行以 States.Timeout 失败，而不是无限期挂起。如果你的 agent 发送心跳，也设置 HeartbeatSeconds 以更快发现死掉的 agent。捕获错误并路由到失败或人工审核路径。 重试时使用稳定的会话 ID。 把 sessionId 设为从执行上下文派生的值（比如 Step Functions 执行名），这样重试会恢复同一个 agent 会话，而不是重新开始。task-token 模式里在派发函数中设置；直接集成模式里在 Task 状态参数中设置。 打开 AWS X-Ray。 在 Step Functions 和 Lambda 配置中启用 Tracing: Active。X-Ray 会精确显示 agent 思考了多久、调用方等待了多久，确认你的模式确实在间隔期间释放了计算。 按速度而不是按 agent 的负载来配派发函数。 派发函数只序列化请求并调用端点。256 MB 内存和 30 秒超时通常就够。重活都在 agent 那边。 成本 这些模式会产生 Lambda 计算、Step Functions 状态转换、durable function 执行存储的费用，以及所有方案共有的 Amazon Bedrock AgentCore runtime 和 Amazon Bedrock 模型推理费用。agent 和模型成本在四种情况下相同。架构只改变编排开销，以及阻塞反模式中浪费的空闲 Lambda 计算。当前价格请查看各服务的定价页。 结论 把 AI agent 放进流水线是简单的部分。让它被经济地调用，才是区分原型与生产设计的地方。阻塞在 agent 上的 Lambda 函数实现简单，却在悄悄烧钱——计费时间里的大部分都在等待。本文展示了三种规避方式：需要在路径里保留 Lambda 时用 task-token 回调；最简单的情况用直接服务集成；偏好用代码表达编排时用 durable function。三者由同一个 Amazon Bedrock AgentCore agent 驱动，它会根据被调用方式调整自己的响应。 完整示例（完整的 agent、状态机定义和 durable function）在 GitHub 的 sample-bedrock-agentcore-async-stepfunctions 仓库。想深入的话，可以看 Amazon Bedrock AgentCore 关于异步任务处理的文档。生产部署请使用 Amazon Bedrock Guardrails，对 agent 的输入输出实施负责任 AI 控制。 原文： #Bedrock# #Serverless# #StepFunctions#

👀 发个有趣的东西 @predictdotfun 徽章系统大揭秘，昨天领到了第一枚徽章，参与世界杯活动的用户，个人主页的 Badge Showcase 里会亮起一枚 Predict Cup 2026。旁边并排躺着几枚灰色的，点开弹窗，一共是10枚。 那么问题来了剩下这些徽章要怎么拿？ 1⃣ 9 枚成就徽章 清单如下： Run It Up｜Market Wizard｜High Noon｜In The Black｜Ascension｜Committed｜Going The Distance｜Heavy Is The Head｜High Roller 其中 pp 那枚值得留意，pp大概率指 Predict Points。说明至少有一枚徽章直接挂钩积分体系。 2⃣名字能猜出什么 官方还没有公布任何解锁条件，下面的解读只是基于名字的语义联想： Run It Up — 滚仓、连续盈利，把利润往上翻。 Market Wizard — 致敬经典交易书《Market Wizards》，大概率是高胜率或高盈利门槛。 High Noon — 西部片决斗梗，可能与特定时刻下单或 50c 对半盘口有关。 In The Black — 会计术语"账面为正"，累计盈利转正。 Ascension — 等级或排行榜攀升，图标是领奖台。 Committed — 图标带 pp，可能与积分系统有关。 Going The Distance — 拳击术语"打满全场"，持仓到结算不提前平仓。 Heavy Is The Head — 出自莎士比亚"欲戴王冠，必承其重"，排行榜登顶。 High Roller — 赌场术语"豪客"，大额交易量。 3⃣ 剩下的可能还不止9枚 除了弹窗里这些，我在挖掘过程中还发现了一个叫 OG 的徽章存在的痕迹。它出现在某个账号的页面上，但没出现在我自己这里。 官方目前只先点亮了世界杯这一枚，其余全部挂在那里等待激活。这像是准备长期运营的成就系统，期待一下后续发展吧。 #预测市场#

看到好多帖子都提到 @ROCKST4R ，确实不仅人美，中英文都非常棒！ 在 @consensus2026 期间，于 @lagrangedev 和 @0G_labs 的 House of AI ，介绍了完整的 Cysic AI 套件。 我也看到了 @cysic_xyz 团队确实是在做实事，成员也非常的努力！ 我自己也在研究AI 相关的行业，而AI竞争的底层逻辑也在发生改变！ 过去两年，市场讨论的核心始终是谁的参数更大，谁的推理更强! GPT、Gemini、DeepSeek 轮流成为焦点！ 但当模型能力逐渐趋同时，一个更现实的问题开始暴露：绝大多数用户，并不想自己搭建 AI 工作流。 他们不关心 Prompt 工程，也不关心底层推理架构。 真正影响 AI 大规模落地的，从来不是“模型够不够强”，而是：AI 能否像操作系统一样，被普通人直接使用。 而这正是 Cysic AI 2.0 的核心方向! 真正决定 AI 能否进入生产环境的，往往不是模型，而是部署能力。 这也是 CyHost 最关键的意义！ 过去大量 AI 项目停留在 Demo 阶段，本质原因并不是功能不够，而是：部署复杂、运维昂贵、稳定性不足。 而 CyHost 想做的，是把 AI 部署压缩成一句极简逻辑： 30 秒部署， 7x24 小时运行， 统一托管。（Pick it. Pair it. Run it.） 所以，未来开发者不一定会绑定某一个模型，而是谁更快， 谁更便宜， 谁更适合当前任务，系统就调用谁。

回家发现门口好几个快递，拆！！ 终于收到OKX跑步速干衣套装，真好，这下我配速要进5分了？ 其实 $OKB 爆拉也有预言的，不信你看盒子上写的“Dont wait for the moon, run to it.”

感谢 #OKX# 送来的中秋运动套装. @mia_okx 从上个月开始每天早上30分钟有氧热身然后篮球基本功训练，晚上下班后力量训练，配合上山姆健康饮食计划已经瘦了不少了，脂肪也明显下降，肌肉含量提升. 现在已经到了白天不做有氧就没有精神工作，完全代替了咖啡的作用，晚上下班不去做力量就觉得今天有很强的负罪感，不知道是不是心理作用，每次我坚持运动的时候行情就会非常的好. DON'T WAIT ！ FOR THE MOON. RUN TO IT 正心正念 ！坚持长期价值 @star_okx