Turning My AI Development Squad from Goldfish to Command Center: Lessons & Tools
The author recounts how an AI‑driven development pipeline initially seemed promising but quickly ran into costly context‑loss issues, leading to a redesign that introduces a single commanding agent and expert sub‑agents, dramatically reducing token costs and improving workflow efficiency.
Background
I was excited when I first saw the CCPM project on GitHub with 4.7K stars, imagining it as the ultimate answer to agile development by automatically analyzing requirements, creating tasks, and handling parallel development.
Pitfalls
After deploying the system, each sub‑agent suffered from a "short‑term amnesia" lasting only about seven seconds, forcing them to repeatedly reload all project documentation, code standards, and API definitions. This caused massive token consumption because the agents repeatedly fetched the same context.
Investigation revealed that Claude Code sub‑agents do not share context with the main agent, leading to duplicated, chaotic, and expensive context loading.
Root Cause Analysis
The problem was not the AI itself but the management model: a flat committee of equal agents constantly shouting for information, resulting in overload and unclear responsibility.
The solution is a "command‑center" structure with a single Project Manager (PM) agent that holds all write permissions and makes final decisions, while other agents act as domain experts (frontend, backend, testing, etc.) that only produce analysis reports.
Implementation – cc-devflow
Based on this model, I created cc-devflow , a set of commands that streamline the workflow: /flow-new "REQ-123|用户下单支持" This single command launches a new requirement, assigning it to the AI PM (nicknamed "老王"). /flow-status REQ-100 --detailed Provides a transparent progress report showing the status of requirement analysis, task breakdown, coding, and testing. /flow-restart "REQ-100" Allows the workflow to resume from the point of interruption without redoing work.
Results
Using cc-devflow, token costs dropped dramatically because the context is now unified under the single commander agent. I transitioned from a code‑centric role to an AI product‑manager role, focusing on designing prompts and SOPs for the agents.
The project, along with all lessons learned, is open‑sourced on GitHub. https://github.com/Dimon94/cc-devflow You can instantly set up your own AI development team with:
npx tiged Dimon94/cc-devflow/.claude .claudeHow this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.