Tristan opens not with a product pitch, but with a mismatch: the agents you and I are running today cannot touch the one thing that actually moves markets, time.
Tristan 的开场不是推销产品，而是指出一个落差：我们现在跑的智能体，碰不到真正让市场动起来的那件事——时间。
"I'm Tristan, Co-founder and CEO of Elfa AI."
"我是 Tristan，Elfa AI 的联合创始人兼 CEO。"
Elfa AI tracks a massive number of real-time data points globally and converts them into real-time signals. Their clients include traders, exchanges, and news platforms like CoinGecko. Keep that in mind, because the rest of this talk is the story of how that stream of signals ends up plugged directly into an AI agent.
Elfa AI 实时追踪全球大量数据点，并把它们转化为即时讯号。客户包括交易者、交易所，以及 CoinGecko 这类新闻平台。请先把这件事放在脑里，因为这场分享接下来的故事，就是这条讯号流如何被直接接进一个 AI 智能体。
Tristan starts by acknowledging how far we have come. Agents today can build personal dashboards, spin up your own apps, stand in as a junior developer, and act as a virtual employee that runs errands for you. At the workshop, some attendees were already orchestrating ten agents to run their companies.
Tristan 先肯定现在的进展。今天的智能体已经可以搭个人仪表板、帮你做自己的 App、当一个还过得去的开发者替身，也可以当你的虚拟员工替你跑腿。现场就有参与者同时编排十个智能体在运营自己的公司。
Then he pivots.
然后他话锋一转。
"But if you think about the ultimate goal of AI, like AGI — agents (not aliens!) taking over the world and running everything — we're still falling short."
"但如果你想想 AI 的终极目标，比如 AGI——智能体（不是外星人！）接管世界、运转一切——我们还差得远。"
The catch is hidden in what the tasks all have in common. Everything we currently hand off to an agent is non-time-sensitive. If an AI employee replies in an hour, it is fine. If it replies tomorrow, it is also usually fine. Nobody is trusting these agents with jobs where being late by sixty seconds is the whole game.
关键藏在这些任务的共通点里。我们现在交给智能体的事，几乎全是非时间敏感的。AI 员工一小时后才回复，没关系；隔天才回，通常也还好。没有人把「慢六十秒就全盘皆输」的任务交给它们。
Tristan lists three that current agents simply cannot touch:
Tristan 点出三个目前智能体完全碰不到的场景：
"Because agents currently lack the ability to react to real-world data. They're in a closed loop, on your computer, on your server. They have information about you, but they don't know what's been happening in the world."
"因为智能体目前缺少对真实世界数据的反应能力。它们被困在封闭回路里，在你的电脑上、在你的服务器上。它们有关于你的资讯，但不知道外面的世界发生了什幺。"
The picture Tristan draws is of a pilot flying blind: instruments wired only to what you personally typed in, nothing wired to the weather outside the cockpit. The agent is brilliant at reacting to you. It is completely deaf to everything else.
Tristan 给的画面是一个在盲飞的飞行员：仪表只接到你亲手输入的资料，完全没接舱外的天气。智能体对你反应灵敏，却对世界上其他所有事情彻底失聪。
AI agents today are reactive to your commands, but completely blind to the real world.
今天的 AI 智能体对你的指令是反应式的，但对真实世界彻底失明。
"Okay, some of you who've already set up your own agents might be thinking: 'That's not right man! What is this guy talking about? Because my agents can run on cron jobs. They can run every 4 hours, 2 hours, 30 minutes...'"
"好，在座有些已经自己架过智能体的人可能在想：『不对啊兄弟！这家伙在说什幺？我的智能体可以用 cron jobs 啊，每 4 小时、2 小时、30 分钟跑一次……』"
Tristan is sympathetic, and then surgical. Cron jobs look like a real solution, but they collapse into a two-sided trap.
Tristan 先表示理解，然后下刀。Cron jobs 看起来像个真正的解法，实际上会陷进一个两面都输的困局。
Check infrequently and you miss the window. The event happens at minute 17 of an hourly cron, the agent wakes up at minute 60, and by then you are already late. Check frequently and the bill explodes: at five-minute intervals you are paying for 288 checks per day per condition, and the overwhelming majority of those checks find nothing.
检查频率低，就会错过窗口。事件在整点后第 17 分钟发生，智能体在第 60 分钟才醒，那时已经迟了。检查频率高，成本就炸：每 5 分钟一次、每个条件每天 288 次检查，而这里面绝大多数什幺都查不到。
| Approach | Cost Profile | What Actually Happens |
|---|---|---|
| 做法 | 成本样貌 | 实际会发生什幺 |
| Low-frequency cron (every 2–4 hrs) | Cheap | Misses time-sensitive events entirely |
| 低频 cron（每 2–4 小时） | 便宜 | 时间敏感事件直接错过 |
| High-frequency cron (every 5 min) | 288 checks/day/condition, most find nothing | Cost explodes, still not truly real-time |
| 高频 cron（每 5 分钟） | 每条件每天 288 次检查，多数一无所获 | 成本爆炸，而且仍不算真正即时 |
| Event-driven (Elfa Auto) | Pay only when a condition actually fires | Agent wakes exactly when it needs to |
| 事件驱动（Elfa Auto） | 只有条件真的触发才付费 | 智能体在该醒的那一刻才醒 |
"But the point is, what I'm saying is: they can't handle time-sensitive things."
"但重点是，我要说的是：它们处理不了时间敏感的事。"
Tristan's insight is almost embarrassingly simple once you hear it. Monitoring and reasoning are two different jobs. Agents today are being asked to do both at once, and they are bad at the first one.
Tristan 的洞察几乎是听了会觉得「就这样？」的那种简单。监控和推理是两件事。今天的智能体被要求同时做这两件事，而它们其实不擅长第一件。
"The insight is that monitoring and reasoning are two different jobs. An agent shouldn't have to do both."
"关键在于：监控和推理是两件不同的工作。智能体不应该两件都扛。"
Line up what Elfa AI already does with what AI agents are missing, and the shape of the fix becomes obvious. Elfa converts real-time world data into real-time signals. Agents are missing exactly that: real-time signals about the world. You do not need a smarter agent, you need to wire in a perception layer.
把 Elfa AI 在做的事和 AI 智能体缺的东西并排放在一起，解法的形状就自己浮出来了。Elfa 把真实世界的即时数据转成即时讯号；智能体缺的正是：关于世界的即时讯号。你需要的不是更聪明的智能体，而是帮它接上一层感知。
"Elfa Auto is a condition engine and trigger pipeline. You define what you care about, it monitors continuously, and it wakes your agent up when conditions are met."
"Elfa Auto 是一个条件引擎加触发管线。你定义你在乎什幺，它持续监控，条件达到时就把你的智能体叫醒。"
Two components, in plain language. The condition engine sits in front and watches the world 24/7 for the things you asked it to watch. The trigger pipeline sits behind it and hands those events off to whatever agent, webhook, or chat channel should hear about them.
拆开来讲就两块。条件引擎坐在最前面，24 小时盯着你叫它盯的东西；触发管线坐在后面，把这些事件递给该知道的智能体、webhook 或聊天频道。
"Your agent is really smart, you've got AI running on your computer. But it can't reach out to the real world on its own. The idea is to let your agent focus on reasoning, while a separate system handles real-time data monitoring."
"你的智能体其实很聪明，你电脑上跑着 AI。但它自己搆不到真实世界。思路就是让智能体专心推理，实时监控交给另一个独立的系统去处理。"
Monitoring and reasoning are two different jobs. An agent shouldn't have to do both.
监控和推理是两件不同的工作，智能体不应该两件都扛。
"This is what I did two days ago. I connected my agent to Elfa Auto. I basically told it: 'Okay, track the technical analysis for all these tokens, and if multiple indicators tell you hey, maybe you should buy a token right now, let me know.'"
"这是我两天前做的事。我把我的智能体接上 Elfa Auto。我基本上就跟它说：『好，帮我追所有这些代币的技术分析，如果多个指标同时告诉你现在该买某个代币，就通知我。』"
Two nights before the workshop, the notification came in. The agent said the signal was strong. Tristan decided to short. He went to sleep.
Workshop 前两个晚上，通知进来了。智能体说讯号很强。Tristan 决定做空，然后就去睡了。
"And then I woke up with over 30% returns. Pretty nice to make that much on a token I didn't even know the name of!"
"然后我一觉醒来，收益超过 30%。在一个我连名字都叫不出来的代币上赚这幺多，感觉还真不错！"
I woke up with over 30% returns. On a token I didn't even know the name of.
我一醒来，收益超过 30%。而且那个代币我连名字都叫不出来。
The point Tristan is making is not that anyone should copy his trade. The point is the workflow. No manual research. No chart watching. The system surfaced the signal, the agent did the analysis, and the only human moment in the loop was one decision.
Tristan 要讲的不是叫大家跟着做这笔交易。重点是整套工作流：没有人工研究、没有盯盘，系统把讯号推出来、智能体做分析，整个回路里唯一需要人的时刻，只剩下一个决定。
"Or if you're ambitious, you can go even crazier. You can chain multiple actions together."
"或者你比较狂，可以玩得更疯。把多个动作串起来。"
The chained version looks like this. Elfa Auto fires a trigger. The agent runs a secondary analysis on the context: is this really what we think it is, how loud is the signal, do other conditions agree? If the confidence clears a threshold, say 70%, the agent proceeds. Depending on how you wired it up, the agent either pings you on Telegram or WhatsApp, or it goes ahead and executes the trade itself.
链式版大致长这样：Elfa Auto 触发讯号，智能体对上下文做一次二次分析——这真的是我们以为的那件事吗？讯号有多响？其他条件是否一致？当信心度过门槛（比方 70%），智能体就继续走下去。接法不同，智能体可以选择在 Telegram 或 WhatsApp 敲你一下，或者直接自己把单子打出去。
That last path is the interesting one: a fully autonomous loop. You sleep. The system watches, reasons, executes. You wake up to a trade that already happened.
真正有意思的是最后那条路：整个回路完全自主。你睡觉，系统在看、在想、在出手。你醒来，那笔交易已经发生了。
"So this is just using trading as an example, but the possibilities are infinite."
"这只是拿交易举个例子，可能性其实无限多。"
Anywhere an agent needs to react to something happening outside of itself, the same architecture applies. PR crisis management, catching a narrative before it lights up. Sales timing, hitting a prospect inside their high-intent window. Market intelligence, surfacing moves from rivals the moment they land. Risk monitoring, waking a human only when a position has actually crossed a threshold that matters.
任何「智能体需要对自己以外发生的事作出反应」的情境，都可以套这套架构。公关危机管理，在叙事点燃前先接住它；销售时机，在潜在客户仍在高意向窗口内时打过去；市场情报，对手动作一落地就浮上桌；风险监控，只有部位真的越过门槛才惊动人类。
"If you want to catch any unusual volume spikes on exchanges and look into it further, you can set up that kind of notification and let your agent work while you sleep."
"如果你想在交易所上捕捉任何异常成交量飙升、再进一步去看一看，你就可以设这种通知，让智能体在你睡觉时替你工作。"
"Soon, next month, we'll have much more data. 40+ data sources. We're known for tracking apps, Telegram, all kinds of technical data sources. And we actually batch-process them almost every few minutes."
"很快，下个月我们会有多得多的资料。40+ 资料源。我们本来就以追踪 App、Telegram、各种技术资料源闻名。而且我们实际上几乎每隔几分钟就批量处理一轮。"
Concretely that means app data, Telegram, and a widening menu of technical feeds, all batched and refreshed on the order of minutes, not hours. The envelope of what a trigger can watch is about to get a lot bigger.
具体来说就是 App 资料、Telegram，以及越来越长的技术资料清单，全部以分钟而不是小时为单位批量刷新。触发器能盯的范围即将扩大一大圈。
Midway through the talk, Tristan hands the mic to his colleague Mingyang. The reason is deliberate. He wants an example that is not about markets at all.
讲到一半，Tristan 把麦克风递给同事 Mingyang。这个安排是刻意的，他想给一个完全不是在讲市场的例子。
"Meanwhile, I want to briefly hand the floor to my colleague Mingyang, because for the feature we're building next month, we've actually already built a POC (proof of concept) that's helped us get pretty solid revenue leads."
"顺便，我想把发言权简短交给我同事 Mingyang，因为我们下个月要上的功能，我们其实已经做了一个 POC，帮我们拿到了不错的营收线索。"
"Hi, I'm Mingyang, Commercial Lead at Elfa. What we've built using Elfa Auto is a lead generation signal pipeline. What does that mean? Basically it means: when a lead gets hot, let me know."
"大家好，我是 Mingyang，Elfa 的 Commercial Lead。我们用 Elfa Auto 搭的是一条潜在客户生成讯号管线。什幺意思？简单说：线索一变热，就通知我。"
Some context. On the B2B side, Elfa ships a lightweight widget that exchanges can integrate with a single line of code. It gives the exchange's traders access to Elfa's AI and the latest insights, with personal positions and platform context folded in.
先给一点背景。在 B2B 这一侧，Elfa 有一个轻量级 widget，交易所只要接一行代码就能集成，让交易所的用户直接用到 Elfa 的 AI 和最新洞察，而且会把个人持仓和平台上下文一并考虑进去。
"Yeah, so just to show more use cases. Because so far I've only covered 3 use cases. He (Mingyang) built this entirely on his own. We didn't even... this isn't an official company project."
"是的，就是想再多展示几个用例。到目前为止我只讲了 3 个。这个是他（Mingyang）完全自己搭的。我们甚至没有……这不是公司的正式专案。"
2 days. 50+ leads. Built by one person, on top of Elfa Auto and Elfa AI.
2 天。50+ 线索。一个人，在 Elfa Auto + Elfa AI 上面做出来的。
If you want to wire your own agent to a real-time monitoring system, paste this prompt into whatever agent you are already running — Claude, OpenClaw, anything with web access:
如果你想把自己的智能体接到一套即时监控系统，把下面这段提示词贴给你现在已经在用的智能体——Claude、OpenClaw，或任何有网路访问能力的都可以：
I want to use auto by elfa, which is a reliable condition engine + trigger pipeline and use it via x402. You can find the latest documentation here: https://docs.elfa.ai/auto/overview.
I want to use auto by elfa, which is a reliable condition engine + trigger pipeline and use it via x402. You can find the latest documentation here: https://docs.elfa.ai/auto/overview.
Then add your actual request — for example, "notify me whenever there is an unusual volume or price change in BTC / SOL / ETH / HYPE in the past 1 hour that may indicate an impending dump or upcoming rally." Your agent will read the docs, set up the conditions, and start monitoring. You can also set things up manually at elfa.ai. Documentation lives at docs.elfa.ai.
接着把你真正的请求加上去——例如「过去 1 小时内 BTC / SOL / ETH / HYPE 出现异常成交量或价格变化、可能暗示即将下跌或启动行情时，请通知我」。你的智能体会自己把文档读完、把条件设起来、开始监控。你也可以到 elfa.ai 手动设定，文档在 docs.elfa.ai。
"Very cheap, about 4 cents per query."
"非常便宜，大约每次查询 4 美分。"
You will need a wallet to pay for queries; cost sits around $0.04 per query. That is the part Tristan wants you to remember when you compare the architecture to a high-frequency cron: you are not paying for the checks, you are paying for the events.
你会需要一个钱包来付查询费用，大约是每次 $0.04。这也是 Tristan 希望你在跟高频 cron 做比较时记得的那一点：你付的钱不是花在「检查」上，而是花在「事件」上。