Agents Join the Group Chat to Work, While Humans Stay in the Review Seat: Teamily’s 32x Growth and a Community Maintenance Experiment

Teamily AI released a set of growth data: in one quarter, registered users went from 10,000 to 320,000—32x. More concrete than the numbers is that a user used it to run a community-maintenance experiment.

In the WeChat group, Scout organizes new issues and PRs by priority, Reviewer does the first round of code review, and Digest drafts bilingual Chinese-English project weekly reports and organizes weekly meeting notes. Loop strings the three into a Scout → Reviewer → human review pipeline. Most things ultimately land with a human to click confirm.

The Scout experience was described as great: it detected more than 200 inactive users from the WeChat group in one go. But when it came time to remove them, the user still had to do it manually. One reply complained bluntly: if 200 inactive users are detected but still have to be kicked manually, how is that different from WeChat’s built-in features? The complaint was blunt, but it hit the boundary of collaboration.

Reviewer tested with already-merged PR #3338 and briefly compared it with the earlier review. It found no new bugs, but for specific code paths it offered some code details worth refactoring. There’s a methodological issue with the test here. A reply pointed out that when using an already-merged PR to test Reviewer, it’s best to provide only the changes and the context visible at that time, while hiding later fixes, human comments, and the final conclusion. Otherwise, it’s easy to mistake restating historical answers for the agent’s review ability.

The biggest difference from most chatbots is that the agents in the group share the same context—they even read more than a month of discussion in the WeChat group. This detail is crucial. Another question is more practical: after sharing a month of chat history, how does the agent find the decisions that are truly relevant now without bringing up outdated conclusions? Someone also asked how Scout handles privacy and security when it reads the WeChat group directly.

Teamily AI positions itself as a social network for humans and agents. Version 2.0 emphasizes real-time human-agent collaboration, unified context and Living Memory, custom agent teams, and multi-model routing. Official figures from the company’s post: 100M+ messages, about 35% generated by AI agents; 20K DAU; several thousand paying users; 10K company teams connecting colleagues; in-house LLM and self-hosted inference with 30% higher throughput than public vLLM/SGLang benchmarks; and positive gross margin for paying users.

Enterprise services include fine-grained access control, private knowledge bases, and private model training and deployment. They also mention hybrid workflows in which local agents handle sensitive data while cloud agents use large models for analysis and content generation. In the growth data, SimilarWeb traffic has already surpassed Town.com, Viktor.com, wabi.ai, and cue.im—products that reportedly have raised $100M to $400M and reached billion-dollar valuations. The company uses this to illustrate its competitive momentum.

Version 3.0 is preparing to build an ecosystem: agentic commerce and wallets, a creation studio and plugins, an AI-native data hub, agent mini apps and an agent network, business application connectors, app-less distribution, and an open model and infrastructure ecosystem. The vision is the Agentic Internet, with AaaS as easy as adding a friend.

Putting agents into a group chat can compress actions such as triage, initial review, and drafting. Responsibility still remains with humans: someone has to click through 200 inactive users, judge code-review conclusions, and set rules for privacy and permissions. Shared context makes agents feel like teammates, but it also brings outdated information, permission boundaries, and historical answers into their judgment.

A user shared Teamily invite codes and a SGLang Omni group chat link. The group owner pays, so using agents in the group is covered by the owner. Those who want to try it themselves can join here: https://teamily.ai/i/6GLFQF and https://teamily.ai/invite/group/4270696980.