After months of practice, I've gradually formed a collaboration pattern between ChatGPT and local agents.
What ChatGPT excels at:
- Requirement discussion and solution design
- Comparing and weighing multiple approaches
- Concept explanation and knowledge queries
- Code review and logical reasoning
What local agents excel at:
- Executing modifications in the actual codebase
- Batch file operations
- Running tests and verification
- Generating structured diff reports
My workflow:
- Clarify requirements in ChatGPT, discuss multiple approaches
- After determining the final approach, describe the task in structured format
- Hand the task to a local agent for execution
- Review the agent's diff and verification reports
- Decide whether to accept the changes
Throughout this process, the human remains the decision-maker. AI is a tool, not a replacement.
Failure experiences are equally important. Some tasks I overestimated the agent's understanding, leading to results deviating from expectations. Some tasks I didn't define boundaries clearly enough, causing the agent to modify files it shouldn't have.
These failures made me realize: the core of AI agent collaboration isn't making AI do more, but making humans think more clearly.