The personal agent was the wrong unit
Every major AI company now ships an always-on agent for the individual user. Teamily AI's founders think that solves the wrong problem. Aiden Chaoyang He, co-founder and CEO, describes the end state he is building toward as an agentic society, in which people and agents work side by side — and he expects agent-based services to gradually replace conventional software-as-a-service over the next five to ten years. The question is not whether your chatbot gets smarter. It is whether your team gets an agent that can sit in the same conversation as everyone else and do real work there.
Teamily AI, which publicly launched in July 2026, is the human-agent platform built on that idea: an AI-native messenger where humans and AI agents coexist as members of the same groups, communities and teams. Instead of opening a separate chatbot, you build your own agents — a researcher, a slide-maker, a finance helper — and drop them into the groups where the work already happens. You mention them, hand them tasks, and they work next to the humans, with shared context and memory across the conversation.
Four products, one direction
Teamily AI is the fourth product in a four-year journey, and that history matters to the story. He and co-founder Salman Avestimehr, Dean's Professor of Electrical and Computer Engineering and Computer Science at the University of Southern California, an IEEE Fellow and PECASE recipient, have collaborated for more than six years, beginning in USC research labs. Their earlier products were FedML, a distributed machine-learning platform, then TensorOpera AI, an agentic AI infrastructure platform, then AgentOpera — each one moving a layer closer to the application, from training and serving models to agents that do work for teams.
That infrastructure background shows up in how Teamily is built. The company trains its own large language models and self-hosts inference, building on the founders' work at FedML and TensorOpera. Using real-traffic monitoring, its serving system claims 30 percent higher throughput than publicly reported vLLM and SGLang benchmarks, while maintaining positive gross margins per paying user. For a product whose pitch is always-on agents inside every team conversation, serving cost is not a footnote — it is the difference between a demo that scales and a business that does. He previously worked on infrastructure for distributed machine learning and large language model pre-training at Meta and Tencent, and the wider team brings researchers and engineers from Meta, Tencent, TikTok, Amazon, Google, Microsoft and Qualcomm.
32-fold in one quarter
On 3 November 2026, three months after the public launch, Teamily reported the kind of quarter that makes a thesis hard to ignore. Sign-ups grew from 10,000 to 320,000 in a single quarter — a 32-fold increase, working out to a compound monthly growth rate of 217.5 percent. The platform reported 20,000 daily active users, more than 100 million messages processed, roughly 35 percent of them generated by AI agents, and several thousand paying customers. The growth, the company said, was almost entirely product-led.
Read those numbers carefully and they tell a second story beneath the headline. A messenger with 320,000 sign-ups and 20,000 daily actives is still early — most sign-ups have not yet become daily habit. But the 100 million messages, with more than a third written by agents, suggest something rarer: in the teams that did adopt it, agents are not a novelty feature tried once. They are producing a large share of the conversation itself. That is the compounding the founders talk about: connected context and living memory mean the tenth conversation with an agent team is better than the first, because the agents remember what the team decided, what it built, and what it still owes.
The product itself is a full productivity stack wrapped around that idea: first-class agents, connected context, living memory, and seven core capabilities out of the box, turning everyday conversations into finished work — websites, email, documents, presentations, video and images, produced inside the tools and channels where the work already happens. Teamily's own framing of the long-term vision is deliberately large: just as Google democratized access to information and services in the web era, Teamily aims to democratize access to intelligence and services in the agentic era, as the central platform of a next-generation Agentic Internet where services are delivered by agents rather than wrapped inside traditional apps.
What has to be true next
What comes next, according to the November announcement, is expansion from the United States into Latin America and East Asia, including China, Japan and Korea; continued post-training of specialized models on proprietary, permissioned collaboration and agent-task data; and platform connections for agents, models and infrastructure from users, partners and third-party developers. Each of those is a bet that the team-agent pattern travels — across languages, work cultures and app ecosystems — and that the data generated by human-agent collaboration becomes the training set for the next generation of models.
Whether that bet pays off is still an open question, and the honest version of Teamily's story includes its open questions. The company has not disclosed revenue, and a freemium model spanning subscriptions reported from $5 to $1,000 a month across global markets still has to prove it can convert hundreds of thousands of sign-ups into a durable paying base. The GeekWire account of the launch put Teamily at 200,000 users in an earlier window, while the broader TensorOpera stack had served more than five million users worldwide — a reminder that Teamily's own user count and its parent platform's reach are different numbers, and should not be conflated.
But the core claim has now survived its first real test. In July 2026, Teamily opened its doors with a thesis: the next frontier is not making individual models more intelligent, but unlocking what humans and AI agents can accomplish together as one team. Three months later, 320,000 people had signed up to try. The team agent, it turns out, was an idea whose time had come — not because agents got smarter in a quarter, but because someone finally put them in the room where the team already works.
The playbook
- Put the agent where the team already is — The personal chatbot was the wrong unit of analysis. Work happens in teams, so agents should join the group conversation where the work already lives, not wait in a separate app.
- Make memory the moat — Connected context and living memory turn conversations into finished work. Without shared memory, every agent session starts from zero and nothing compounds.
- Own the serving layer — Teamily trains its own models and self-hosts inference, building on four years of infrastructure work, so the cost of each agent interaction stays under its own control.
- Let the product do the distribution — The growth was almost entirely product-led: a messenger people invite their team into spreads the way messengers always have — one group at a time.