Coming Soon

We are building Endue. There is nothing to show on screen yet, so instead of screenshots, here is where we are looking. This is not a product announcement. It is a record of why we are building this at all.

How far agent services have come

AI agent services have grown fast over the past two years. First they were chat windows. You typed a question, an answer came back, and that was the whole story. Then models picked up tools. They search, run code, read files. Now they run loops: given a goal, they plan, choose tools, look at the result, and decide the next step.

That progress will continue. Models will hold focus longer and handle tools with more precision, and we have no doubt about the direction. One thing keeps catching our eye, though. Most services today run agents one at a time, inside one window, only while a person is watching. Model capability grows every year, while the vessel that holds agents is still a single chat box. The vessel is what we are building.

From one to many: multi-agent

Human work does not finish alone. A company has roles, and each role carries its own permissions and responsibilities. Agents are heading the same way. One agent that researches well, one that edits prose, one that talks to customers. Stuff every instruction into a single do-everything agent and the instructions start colliding, and context runs out fast. Divide the roles and each agent gets simpler while the whole gets more capable.

The moment you run several, new questions appear. Someone has to know who this agent is, what it may do, and what it did yesterday. So the center of a multi-agent system is management: identity, permissions, records. When you hire a person you create an account and grant access first. Agents need the same thing.

More channels change the work

As long as using an agent means opening yet another app, the agent stays a tool. Work happens in KakaoTalk, in Slack, in the inbox. For an agent to become a colleague, it has to go where people already talk: invited into the team channel, answering customer inquiries, picking up the mail thread.

More channels also change what an agent is. An agent inside a chat window only has to answer what it is asked. An agent on a channel never knows when the next message lands. It arrives at night, and ten arrive at once. Always on call, handling things concurrently, with a different voice and different permissions per channel. That is not a prompt problem. That is a systems problem.

How agents divide the work

Once several agents live in one organization, the next question is collaboration. The research agent hands its findings to the writing agent, who drafts; the review agent reads it and sends it back. The flow a human team runs today, agent teams will run too.

Collaboration needs protocol. What travels along when work is handed over, who receives and checks the result, where it returns when it fails. And all of it has to leave a trail. If you cannot trace which agent did what under which authority, that collaboration can never enter an organization. We see agent-to-agent collaboration as workflow rather than an extension of chat. Handoff, review, approval, and audit belong in the system layer.

Toward full autonomy

The last stage is an agent that moves without being told. One that reads the metrics every morning and speaks up when something looks off. One that starts triaging on its own when inquiries pile up. It receives goals rather than instructions, and wakes itself when the time comes.

Autonomy is not neglect. The more autonomy grows, the more precise control has to become. You should see what it is doing in real time, step in at any moment, and stop it instantly. In front of a risky action it should stop and wait for a human to approve. A car with good brakes is the one that can drive fast. Autonomous agents are the same: you can extend autonomy exactly as far as observation, intervention, and stopping reach.

Why we believe we can do this well

You may have noticed what everything above has in common. Identity, permissions, real-time observation, records, channel integrations. None of these are model problems. They are system problems, and systems like these are what we have been building: authentication and authorization, real-time pipelines, messaging-platform integrations.

Our approach starts there too. Governing an agent by prompt has a ceiling. A prompt is a request, and requests can be ignored. Permissions, approvals, and stops must be enforced in the system layer. Building the layer that holds still no matter how the models change is what we do well, and it is the spine of this product.

Why cloud

An autonomous agent must not turn off. To catch the customer message that arrives at 3 a.m., to wake itself at 9 every morning, it has to keep running somewhere regardless of whether a laptop lid is open. Channel integrations ask for the same: receiving a messenger webhook requires an address that is always reachable.

The security boundary is also clearer in the cloud. Credentials live in one place instead of being scattered across personal machines, audit logs accumulate in one place, and when something goes wrong you cut it off in one place. When ten agents become a hundred, a personal device cannot carry that weight. Cloud is less a choice than a precondition of autonomous agents.

The next owner of internet traffic

Close to half of internet traffic is already measured as non-human. Crawlers and bots have held that share for years, and agents are now being added on top: traffic that searches on a person’s behalf, books on a person’s behalf, and talks agent-to-agent over APIs.

That curve only goes up for a while. One person will run several agents, and a single agent generates dozens of calls. In that world, what is needed is a place where that traffic can be created, seen, and governed. When agents become major actors on the internet, those actors need an operational home. That home is what we are building.

An ecosystem completes it

An agent alone has limited use. Usefulness comes from connection: connectors that reach mail, calendars, and document tools, skills that carry ways of working, and conventions that let agents find and call each other. Only when these accumulate does an agent enter real work.

And an ecosystem cannot be built alone. Our plan is to let anyone add connectors, share skills, and link their agents to other agents. The value of a platform, we believe, is the sum of what others build on top of it.

So, coming soon

Multi-agent, multi-channel, collaboration, autonomy. We know the road is not short. We will walk it in order: first a place to build and manage agents, then channels, then collaboration and autonomy. As each part becomes ready, this page is where we will show it.

See you soon.