ByteDance's Doubao Work: cloud agent writes 37k-word research reports and builds its own skills
AI寒武纪 · wechat · 2026-08-28
ByteDance has consolidated TRAE, Coze and Feishu (Lark) under the Doubao brand and launched Doubao Work, a standalone agent for work scenarios, with a free one-month membership at launch. The author tested two cases:
- Cloud deep research: dispatched from his phone in a car — the agent role-played a hard-tech VC chief analyst, searched Tesla Optimus, Figure AI and Unitree progress plus supply-chain cost data, and delivered in about a meal's time a 37,000-word HTML research report and a 10-slide PPT with speaker notes. Notably, it auto-invoked the Feishu lark-slides-pro Skill and read its spec docs before building slides. The report's sharp conclusion: only 38 industry grinding machines exist as a physical ceiling, making 2026 a "hardware mass-production year" rather than a "commercial deployment year," with economically meaningful scale unlikely before 2028.
- Custom skill creation: given just a link to the Seedance 2.5 storyboard prompt manual, one instruction made it read the manual, build a multi-turn-guided skill, generate three 15-second videos (retrying itself after an error), and embed prompts plus videos into a new Feishu doc.
His takeaway: Doubao Work's edge is knowing which tool to call and where output should land; deep Feishu integration embeds it in team workflows. Caveat: verify AI-generated data before real investment decisions.
Related event: ByteDance Unveils "Doubao Work" Office Agent(3 posts)→
More from coding & agent
- AQuA paper: if an agent can edit its tool schema, you're not evaluating the same system — creditme7 · 2026-08-28
- 4 Hours on a Train: A Fold Plus Remote Desktop Is Enough to Run All Your Agents — kevinkern · 2026-08-28
- Accio Open-Sources CommerceAgentBench for Real Agent Tasks — kimmonismus · 2026-08-28
- Code World Model: A Coding Agent as World Brain, Code for Rules, Video for Rendering — burny_tech · 2026-08-28
- MIT-licensed tools reduce coding agent context waste without secondary LLMs — bonsaisushi · 2026-08-28
- Testing IRL evaluation method saves 20-70% compute cost — xeophon · 2026-08-28