The colleague you @ in your work group chat might not be human for much longer.

On July 2, Kunlun Tech released the Skywork 3.2 LLM upgrade, whose core feature is Skywork Tags — issuing AI Agents an “employee badge,” letting them join enterprise work chats as formal “digital employees” working alongside humans. This isn't a chatbot; it's an Agent with identity, role, and task boundaries that can be @mentioned, assigned tasks, and performance-tracked. The same day, AReaL 2.0 was released open-source, providing reinforcement learning infrastructure for self-evolving Agents — letting Agents grow stronger through use.

From Tool to Colleague: The Agent Identity Evolution

Skywork Tags is fundamentally an Agent identity system. Before this, most enterprise AI use followed a “open a dialog box and ask a question” pattern — AI was a tool, humans were users, and you walked away when done. Skywork Tags changes that relationship: Agents get employee IDs, roles (like “data analyst” or “customer service assistant”), and permission boundaries; they can exist in Feishu/DingTalk/WeCom group chats, @mentioned, assigned tasks, and reporting work like real employees.

An ifanr report on Feishu's AI assistant shows a similar trend: from “opening an AI dialog” to “@mentioning AI directly in the group”, interaction shifts from “humans seeking out AI” to “AI on standby within the workflow.” DingTalk's A1 recording card was also selected for Gartner's AI wearables report, indicating the convergence of AI hardware and work scenarios is accelerating.

This shift may look like just an interaction change, but it's actually a fundamental shift in how enterprises perceive Agents: AI is no longer a “fancy search engine,” but a “virtual employee” who can carry specific responsibilities. This means Agents don't just need stronger model capabilities — they need a full stack of “human resources management” infrastructure for identity management, permission control, task tracking, and performance evaluation.

Making Agents Stronger Over Time: RL Infrastructure Fills In

But a “digital employee” stuck at day-one proficiency forever provides no value. That's the problem AReaL 2.0 solves.

AReaL 2.0 is an open-source reinforcement learning (RL) infrastructure for self-evolving agents. Simply put, it lets Agents continuously learn and self-optimize through interaction with their environment, rather than relying on manual annotation and retraining each time. Traditional LLM training is “one-shot” — model capability is fixed after training; AReaL's goal is for Agents to keep evolving during use, like a new hire going from onboarding to proficiency.

InfoQ's analysis notes that a major pain point for Agent deployment at scale is “getting dumber with use” — with fixed model weights, Agents can't learn from successes and failures and repeat the same mistakes. AReaL 2.0's RL infrastructure lets Agents learn from environmental feedback and gradually optimize strategies. This is especially important for enterprise scenarios: a customer service Agent that requires human intervention for every new issue isn't automation.

These two stories together outline the next stage of Agent development: from “usable” to “good” to “gets better over time”. Skywork Tags solves the Agent's “entering the workplace” identity problem; AReaL 2.0 solves the Agent's “continuous growth” capability problem. When Agents can have identities, roles, learning, and improvement just like humans, they truly evolve from tools to colleagues.

Of course, Agent “onboarding” also raises new questions: who manages these digital employees? How is their performance evaluated? Who's responsible when they make mistakes? These questions don't have answers yet, but the direction is clear — in future work groups, humans and AI will be mixed together, and you might not be able to tell them apart.