A Quiet Beta With Big Ambitions
On the evening of 18 April 2025, ByteDance slipped an invite‑only beta of a new product called Coze Space into its fast‑growing Coze developer community. Officially described as a “general agent platform”, Coze Space aims to evolve the original Coze no‑code bot builder into an end‑to‑end workspace where multiple AI agents can think, plan, and execute complex tasks for you — from drafting a five‑day Japanese travel guide to crunching Python for equity research.
The timing is no accident. Just 24 hours earlier ByteDance’s cloud arm, Volcano Engine, introduced an OS Agentsolution and a new “deep‑thinking” 200‑billion‑parameter model inside its Doubao family, signalling a company‑wide push toward agentic computing.
From “Answering” to “Doing”
Unlike classic chatbots that reply in text, Coze Space lets an agent:
- Understand the goal → break it into sub‑tasks → call the right tools → deliver structured output (HTML, PPT, Feishu docs, audio, images).
- Choose between Explore mode (rapid, best‑effort execution) or Plan mode (think‑then‑act for high‑stakes jobs).
- Tap an early catalogue of Expert Agents such as “User‑Research Analyst” and “Huatai A‑Share Insight” that bundle proprietary data and Python pipelines behind a single prompt.
For developers, everything that already exists in Coze — prompt cards, plug‑ins, knowledge bases, long‑/short‑term memory, multi‑model switching (Doubao, Tongyi‑Qianwen, Kimi, MiniMax, Zhipu, Baichuan, etc.) — remains available as building blocks.
A Standard Plug‑In Pipe: MCP
Coze Space implements the emerging Model Compatibility Protocol (MCP), a vendor‑neutral “USB‑C for agents”. The beta bundles adapters for Feishu Sheets, Amap navigation, TTS, image generation, and more; third‑party developers will soon be able to ship their own tools through the same interface.
Growth Hack—Invite Five, Get Five
ByteDance is using a classic viral loop: each new user receives five invite codes, and completing your first task grants five more. The company insists this is “not artificial scarcity” but a way to stress‑test capacity while expanding organically.
Early Hands‑On: Promising—but Not Perfect
Tech site GeekPark spent a night with Coze Space. In Explore mode the agent produced a detailed Kansai–Kumamoto itinerary (maps, budgets, Japanese phrases) in about 15 minutes and generated outfit collages for a week‑long weather forecast. Yet an equity‑analysis task stalled for hours when a Python script hit API‑permission errors.
The takeaway: ByteDance’s first public agent is already useful for consumer‑grade planning and content generation, but enterprise‑class reliability will need polishing.
Top AI Product 
Leave a comment