The 23-year-old founding father of Emagen AI argues your entire agent {industry} is optimising the flawed unit. His reply is an working system the place AI drives the work and calls on people, not the opposite means round.
Each week, one other AI agent startup launches. They write code, draft emails, generate slides, analyse knowledge. Every one guarantees to make you extra productive. Yimao Zhou thinks they’re all fixing the flawed downside.
Zhou is the founder and CEO of Emagen AI, the corporate behind Cagen, what he calls an “OS Stage Agent,” an organizational working system powered by AI. Backed by MiraclePlus (previously YC China) and its legendary founder Qi Lu, Zhou is betting that the way forward for AI isn’t about making people sooner. It’s about making groups essentially totally different. Right here’s a Q&A with the founder to know what which means, and why he thinks 90% of at this time’s AI agent firms received’t exist in three years.
You’ve mentioned that AI is definitely making groups worse. That’s a reasonably contrarian take given that each AI firm is promising productiveness positive factors. What do you imply?
Take into consideration what occurs whenever you give each particular person on a five-person staff their very own AI assistant. Every particular person produces extra, sooner. The product supervisor generates specs sooner. The engineer writes code sooner. The designer iterates sooner. Sounds nice, proper?
However right here’s what truly occurs: the output diverges. Everybody’s transferring sooner in barely totally different instructions, and no one notices till it’s too late. The bottleneck in a staff was by no means “one particular person works too slowly.” It was at all times “are these 5 individuals constructing the identical factor?” AI instruments speed up the elements that weren’t bottlenecks and make the true constraint — coordination — worse.
60% of information employees’ time goes to what I name coordination prices — syncing progress, writing standing updates, relaying data between individuals, ready for approvals. And these prices don’t simply exist between people. Within the AI period, they multiply: human-to-agent coordination, agent-to-agent coordination, the overhead of retaining everybody and every part on the identical web page. AI is optimizing the opposite 40%, the precise doing, and fully ignoring the 60%. That’s not only a missed alternative. It’s a directional error.
So what ought to the {industry} be constructing as a substitute?
Each main computing shift follows the identical path: instruments come first, then platforms, then an working system emerges. PCs had standalone software program earlier than Home windows. Cellular had particular person apps earlier than iOS and Android unified the expertise. Cloud had scattered providers earlier than AWS grew to become the infrastructure layer.
AI is on the identical curve. Proper now we’re within the “standalone instruments” section. Lots of of brokers, every doing one factor nicely, none of them speaking to one another. The platform section is simply beginning. The OS section hasn’t occurred but.
That’s what Cagen is. Not one other AI device. The working system layer for the way organizations work with AI.
“OS Stage Agent” is a giant declare. In concrete phrases, what does that really appear like?
Right here’s a structural downside no one’s addressing. Notion constructed Notion AI. GitHub constructed Copilot. Salesforce constructed Einstein. Each SaaS firm is embedding AI, however their incentive is to make their very own product stickier, to not join you throughout instruments. Notion AI makes Notion extra priceless. It has zero incentive that will help you bridge Notion to GitHub to Linear to Slack.
Meaning cross-tool intelligence is structurally inconceivable for any incumbent to construct. It will possibly solely come from an unbiased layer.
Now, some individuals will say: “What about MCP? Anthropic’s Mannequin Context Protocol already connects AI brokers to a number of instruments.” True, and MCP is nice. However MCP is a connector protocol. It’s USB, not an working system. It lets one particular person’s agent plug into that particular person’s instruments. There’s nonetheless no shared organizational context, no persistent staff reminiscence, no cross-role orchestration. MCP truly advantages us. The extra standardized the plumbing will get, the simpler it’s to construct an OS on prime.
Cagen is that OS. However right here’s what actually separates it from every part else, and that is the half most individuals miss. Each AI product at this time, together with those that decision themselves “staff AI,” works the identical means: seize data, arrange it, and look forward to a human to question it. The human remains to be the driving force. The AI is a librarian.
Cagen inverts that. Our brokers have objectives and context. They constantly cause about what must occur subsequent primarily based on the staff’s targets, the mission state, and organizational context. After they want human judgment — a call, an approval, inventive enter — they name on the human. The human is a useful resource within the system, not the operator of the system.
That’s what makes it OS-level. An working system doesn’t wait so that you can manually handle each course of. It runs, it schedules, it handles occasions. It calls on you when it wants you. That’s how Cagen works for groups — groups of people and AI brokers working collectively.
The AI market is brutally aggressive. Traders will ask: what’s your moat?
After six months of utilizing Cagen, what makes it irreplaceable isn’t any function we constructed. It’s what your staff constructed on prime of it: determination patterns, communication habits, high quality requirements, workflow information. All of that’s deeply coupled to your particular organisation. A competitor can clone each function of Cagen. They can’t clone six months of your staff’s amassed intelligence.
This is identical cause Salesforce has industry-leading retention. It’s not as a result of the CRM is irreplaceable. It’s as a result of the information, processes, and automations working on it are irreplaceable. The product turns into an organisational asset, not a software program subscription.
However right here’s the vital distinction: that stickiness comes from amassed worth, not synthetic lock-in. We’re not trapping anybody. Groups keep as a result of they don’t wish to lose what they’ve constructed.
Particular person AI reminiscence is nicely understood. How is organisational reminiscence totally different?
Basically totally different. Particular person AI reminiscence scales linearly. I study one thing, I profit. Organisational AI reminiscence has community results. One particular person’s studying advantages everybody on the staff — and each agent on the staff. The compounding fee is n-squared, not n.
That’s why a “Workforce Agent” isn’t only a multiplayer model of a private agent. It’s a totally totally different species. When one staff member refines how aggressive evaluation will get executed, that information instantly elevates everybody else’s output — and each agent’s output. When the system learns how your group defines “good,” what high quality appears like, what tone you employ, the way you construction selections, it raises the ground for each piece of labor throughout the corporate, whether or not it’s executed by a human or an agent.
Private AI makes one particular person higher. An OS Stage Agent makes the organisation smarter as a unit.
You retain saying “staff.” However the development proper now’s the alternative: extra solo founders, extra one-person firms. If groups are shrinking, who wants a staff OS?
That’s precisely the precise query, and the reply truly makes our case stronger.
There are literally two traits occurring concurrently, they usually’re squeezing from either side.
On one finish, organisations are getting bigger and extra complicated. World groups, cross-timezone coordination, regulatory overhead, multi-vendor provide chains. The coordination burden inside giant organisations retains rising.
On the opposite finish, people are getting smaller and extra unbiased. Layoffs are accelerating. The freelancer financial system, digital nomads, solo founders, one-person firms — they’re all exploding. However right here’s what individuals miss: a solo founder doesn’t work alone. They rent a contract designer on Fiverr, a contract developer on Upwork, a fractional CFO, a advertising and marketing advisor. The “staff” nonetheless exists. It’s simply not a set org chart anymore. It’s fluid, momentary, project-based. And more and more, it contains AI brokers as full staff members.
Each ends want the identical factor: an orchestration layer. And that want goes to accentuate. Work is atomizing. You’ll see increasingly granular wants matched with increasingly specialised suppliers — on-demand, globally, in actual time. The outdated mannequin was: rent 5 full-time workers, put them in an workplace, handle them. The brand new mannequin is nearer to Uber for work. Assemble the precise individuals and brokers for the precise job, execute, disband.
However right here’s the issue with that mannequin: coordination prices explode. When your “staff” is a rotating solid of freelancers, contractors, and AI brokers who don’t share context, don’t know one another’s working model, and don’t have shared historical past — the coordination downside we talked about earlier will get ten instances worse.
That’s the place Cagen turns into important. It’s the orchestration layer. It holds the organisational context, the mission historical past, the standard requirements, and it dispatches work to the precise individuals and brokers on the proper time. The solo founder doesn’t must handle anybody. Cagen manages the constellation.
So “staff” doesn’t imply 5 individuals in a Slack channel. It means any group of people and AI brokers collaborating towards a aim. The extra fluid and atomised work turns into, the extra you want an OS to carry all of it collectively.
Who’re your first prospects? I’d assume tech startups.
Truly, no, and that is counterintuitive. Tech firms have already got deeply entrenched toolchains. Slack, Notion, Linear, GitHub. They’re locked in, and the switching value of including an OS layer is highest for groups which have already optimised their current stack.
Our greatest early prospects are organisations with excessive operational complexity however with out deep dedication to any particular device ecosystem. We’re at present deployed with a boutique lodge in Pittsburgh, for instance. A lodge operations staff juggles visitor communication, upkeep coordination, shift scheduling, vendor administration: dozens of handoffs per day throughout a number of roles. The coordination prices are excessive, however they haven’t constructed their workflows round a inflexible SaaS stack.
That’s the candy spot: complicated sufficient to wish an OS, versatile sufficient to undertake one. And if it really works in hospitality, one of the vital operationally dense environments for small groups, it really works wherever.
However hospitality, CPG, logistics: these are all very totally different industries. How do you scale throughout all of them with out turning into a customized consulting store?
That is the query everybody asks, and it’s the precise one. The standard reply is: you rent {industry} specialists, do bespoke integrations, and it doesn’t scale. That’s the consulting lure.
Our reply is totally different. Take into consideration the pipeline from buyer acquisition to deployment: understanding a consumer’s operations, figuring out the place AI suits, constructing the precise workflows. There’s no inherent cause that total course of has to depend on people.
The bottleneck at this time is a mismatch. Non-technical customers don’t perceive what AI can and may’t do. On the similar time, they battle to articulate their very own wants clearly. That’s why each AI integration at this time requires somebody who has each area experience and AI experience, and that mixture is extraordinarily uncommon and costly.
Cagen’s roadmap is to fuse these two collectively contained in the product. Ideally, a consumer simply describes what their staff does everyday, together with their firm’s objectives. The system then mechanically understands, decomposes, and constructs the precise workflows. It’s an automatic consulting and execution layer. The AI doesn’t simply run your workflows; it figures out what your workflows must be.
We’re not there but. No person is. However even on the present stage, the method offers us a structural benefit. And the place full automation isn’t doable at this time, we will route particular wants right into a market: people appearing as builders, just like Upwork or Fiverr, however orchestrated by the system. That turns bespoke integration from a consulting downside right into a platform downside. And platform issues scale.
You have been backed by Qi Lu, who determined to take a position ten minutes right into a thirty-minute pitch. That story’s been informed earlier than. What does it truly imply to you now, trying again?
What it means is that he wasn’t investing in a product. He was investing in a judgment.
Qi Lu spent his profession on the OS layer: Government VP at Microsoft, President and COO at Baidu. When he heard me describe the AI agent panorama as “everybody constructing apps, no one constructing the working system,” he didn’t want a demo. He’d lived by that actual sample earlier than. He knew what occurs when somebody identifies the precise abstraction layer early.
Most AI pitches are “we do X higher than Y.” My pitch was “your entire {industry} is constructing on the flawed layer.” He recognised the distinction instantly. That’s what the ten minutes have been about.
Claude Code surpassed $2.5 billion in annualised income by early 2026, contributing to Anthropic’s $44 billion whole run fee by mid-year. OpenAI Codex has 5 million weekly customers. OpenClaw has over 370,000 GitHub stars, greater than the Linux kernel. Whether or not backed by probably the most highly effective AI labs or the open-source neighborhood, the momentum behind AI brokers is very large. How do you compete with that?
I don’t. As a result of we’re not enjoying the identical recreation.
Take a look at what these merchandise truly are. Claude Code is a terminal agent that helps one developer mass-produce code. Codex is identical factor inside ChatGPT. OpenClaw is an open-source private assistant that runs in your laptop computer. They’re all extraordinary at what they do, and what they do is make one particular person extra productive.
Claude Code even has one thing referred to as “Agent Groups.” Feels like staff collaboration, proper? It’s not. It’s one particular person orchestrating a number of AI situations. There’s no shared context between staff members. No organisational reminiscence. No cross-role coordination. Codex’s “Marketing strategy” is seat administration and billing. It doesn’t change how the product works at a staff stage.
That is precisely my level. One of the best-funded, most proficient AI labs on the planet are all converging on the identical factor: supercharging people. They’re constructing probably the most highly effective apps the world has ever seen. However no one is constructing the OS.
There’s a means to consider this that I discover clarifying. The infrastructure for AI-assisted coding — what some individuals name the “coding harness” — is actually a solved downside. It’s a continent. Claude Code, Copilot, Cursor, Codex: the land has been claimed. However the infrastructure for AI-assisted working — coordinating groups, managing objectives, orchestrating people and brokers collectively — remains to be an unlimited blue ocean. There are a couple of small islands, however no continent. That’s the place we’re constructing.
When your engineer makes use of Claude Code and your product supervisor makes use of OpenClaw, every particular person will get sooner. However the coordination between them — the context, the choices, the handoffs — nonetheless travels by Slack messages and standing conferences and Google Docs that no one reads. The coordination prices are fully untouched.
That’s the hole. It’s not a function hole. It’s a layer hole. And it’s not going to be crammed by Anthropic or OpenAI, as a result of their enterprise mannequin is promoting seats to people. An OS for organisations is a essentially totally different product with a essentially totally different structure.
Final query. Three years from now, what does the AI agent {industry} appear like?
Most of at this time’s AI agent startups will likely be useless. Not as a result of they’re dangerous, however as a result of they’re constructing at a layer that’s about to get commoditised. If you’re basically wrapping a immediate round a basis mannequin and optimising for one vertical, your moat is immediate engineering. That’s not a moat. That’s a sand fort.
The survivors will likely be firms that constructed at a layer the inspiration fashions can’t simply soak up. For vertical brokers, which means deep domain-specific knowledge flywheels. For us, it means the OS layer: the orchestration and organisational intelligence that sits above any single mannequin.
However the true disruption isn’t about which firms survive. It’s about what turns into doable. The minimal viable staff dimension for a critical enterprise is about to break down. Issues that required 50 individuals would require 5 individuals plus an AI working system. That doesn’t simply change how firms work. It adjustments which firms can exist. A large variety of enterprise concepts that didn’t pencil out underneath the outdated mannequin all of a sudden change into viable.
Three years from now, individuals received’t ask “what AI device do you employ.” They’ll ask “what OS is your staff working on.”
Yimao Zhou is the founder and CEO of Emagen AI, the corporate behind Cagen. He beforehand studied drugs at Shanghai Jiao Tong College and cognitive philosophy and philosophy of science. He was the youngest founder in MiraclePlus’s F24 cohort. Be taught extra at cagen.ai.
