DeepSeek resident agents: scheduling gains and the engineering cost
What happened
On November 6, Computing Life analyzed how DeepSeek's resident agents split parallel development work between a project manager and background sub-agents, and criticized the weight of its state machine and permission patches. In one comparison, restricting the project manager to dispatching work cut the slowest task's completion time from 29.5 to 15.7 minutes, but cache token use rose 3.8x. Another change sent follow-up messages straight to background sub-agents, cutting project manager wake-ups by about 30% and cache reads by 27%.
Written by AI from the coverage · updated 34 minutes ago
Coverage
Follow the reports to see the story from different sides.
- Computing Life · Share · Yage给编程 Agent 配项目经理：解析 DeepSeek 常驻代理的调度痛点与工程补丁
DeepSeek 常驻代理用项目经理与后台子代理分工处理并行开发任务，文章分析其调度收益，并批评状态机与权限补丁过重。文中对比实验显示，强制项目经理只派工将最慢任务完成时间从 29.5 分钟压至 15.7 分钟，代价是缓存 token 消耗增至 3.8 倍；跟进消息直达后台子代理则减少约 30% 的项目经理唤醒和 27% 的缓存读取。
Heat over time
Not enough continuous observations to draw a trend yet.