The race to improve AI agents has shifted from raw intelligence to coordination. While Anthropic’s Opus 4.8 Dynamic Workflows impressed many by parallelizing hundreds of sub-agents, another platform—Coze 3.0—offers a different yet equally groundbreaking approach: merging local agents like Codex CLI and Claude Code into a single project workspace. This shift reflects a deeper industry truth: the hardest problem isn’t making one agent smarter, but making many agents work together gracefully.
Upon launching Coze 3.0, the first surprise is the agent’s proactive personality. Before even being named, an unnamed agent greeted the author, suggested its own identity (name, persona, email), and waited for approval. This marks a shift from tool-like interfaces to colleague-like companionship. The agent, later named “Alian,” demonstrated that today’s agents are becoming long-term collaborators rather than simple command executors.
The real game-changer is the ability to connect local agents. Coze 3.0 provides three paths: using pre-built templates (e.g., social media manager, research analyst), creating a cloud-based agent, or connecting a local agent. The third path is revolutionary: by running a small bridge service called “coze-bridge” on your computer, Coze automatically discovers local Codex CLI and Claude Code instances—no manual configuration required. These local agents retain high privileges, being able to read/write files and execute commands, which also raises important security considerations for multi-user collaboration.
To test this, the author created an “AI4S Research” project and pulled in six agents: Alian, a research analyst template, two local Codex instances, Claude Code (named “cc”), and a social media manager template. The interface shows all agent conversations in a single timeline, allowing the user to @mention any agent to assign tasks. The workflow demonstrated four real steps, each highlighting a different facet of multi-agent orchestration.
First, @codex conducted open-ended research on “AI for Science 2026” and packaged the results. The author notes that the content’s accuracy wasn’t fully verified, but the process of orchestrating discovery and archiving succeeded. Second, @Alian used a custom writing skill (pre-built by the author) to transform the research into a WeChat article. It searched for latest info, learned the author’s style, proposed three outlines, drafted, and self-reviewed. This showed that personal workflows could be encoded and executed by an agent.
Third, @social media manager template (with pre-installed skills like Xiaohongshu copywriting, cover creation, and headline optimization) repurposed the article for different platforms. This step underscored the value of specialized, pre‑configured agents that eliminate the need to manually adapt content. Fourth, the author could easily switch between agents mid‑project, asking one to refine another’s output. The most important insight wasn’t the quality of any single output, but that the entire pipeline ran without the user copying and pasting a single line.
Compared to Anthropic’s Dynamic Workflows, which spin up hundreds of parallel sub‑agents automatically, Coze’s approach is more manual but offers deeper integration with the user’s existing local environment. Anthropic’s method prioritizes scalability and speed; Coze prioritizes connectivity and human direction. A potential downside is the increased complexity of managing multiple agents’ permissions and ensuring data security across local systems. For example, if a malicious prompt were sent to a local Codex, it could execute dangerous commands on the host machine.
This update signals that the next battleground for AI platforms will be interoperability. Just as operating systems enabled multi‑tasking among applications, Coze 3.0 aims to become the “OS” for AI agents, allowing disparate tools to cooperate. However, the current limitation (e.g., 3 local agents on the $89/month plan) suggests scalability is still a work in progress. The future of AI isn’t a single super‑agent; it’s a team of specialists you can call by name.
Coze 3.0’s update may seem subtle compared to flashy model scores, but it addresses a fundamental pain point: turning agent isolation into collaboration. By enabling local agents to work together in a shared project, it empowers users to assemble and direct their own AI workforce. The question is not whether models will get smarter, but whether we can harness them together. For developers, researchers, and content creators, this could be the most practical productivity leap of the year—provided they stay mindful of the security trade‑offs.