A Developer’s Guide to Mastering AI Collaboration, Boosting Efficiency, and Becoming an Indispensable AI Team Leader.
The Rise of AI Coding Companions & Key Interaction Styles:
A new generation of AI-powered coding assistants and editors is emerging, each offering unique ways to integrate AI into the development lifecycle. You might have heard of tools like GitHub Copilot, which started by revolutionizing code completion, or newer AI-first editors like Cursor and Windsurf.
As you explore these options, you’ll notice a fundamental difference in how you interact with them and what they can do. This typically falls into two main categories:
Chat-Based & Suggestive AI: The agent doesn’t have full control of your editor. This means you’ll typically need to run commands manually and then copy any error messages from the terminal back to the chat.
Agentic “Auto-Driving” AI: This could involve refactoring code across multiple files, scaffolding new features based on a descriptive brief, and attempting to debug issues.
What is Agentic AI? It refers to artificial intelligence systems that can act autonomously and proactively to achieve specific goals with minimal human intervention.
For instance, Cursor, recognized as a pioneer of Agentic AI within IDEs (code editors), can automatically run unit tests, interpret errors, devise solutions, and apply code updates.
- Understand the higher-level goals described by the developer.
- Plan a sequence of actions to achieve those goals.
- Interact with the codebase across multiple files.
- Utilize tools (like terminal commands or file system operations).
- Iterate and self-correct to some extent.
- Operate with greater autonomy than a simple chat assistant.
Chat-Only AI Editor is like a “Car with Advanced Driver-Assistance and a Smart Co-Pilot”.
Agentic AI Editor is like a “Car with Full Autopilot for Software Development”.
I have been using the Agentic AI editor (Cursor) for a few weeks. Here are some tips:
You are a team leader, not a coder.
Change your coder’s mindset to a team leader’s mindset. Motivate and guide your team to do the work. The productivity will boost when you can master the “relationship” with your AI coders.
In order to manage the relationship, you need to know:
- What are the strengths and weaknesses of your coders?
- How to give instructions to your coders?
- How to control work quality?
- How to make them cooperative when you have more than one team member?
Get Engaged, but Avoid Micromanaging
Like managing engineering teams working on projects. As a manager, you need to balance team autonomy and your engagement. Understand what the team does, and give instructions at the right time.
I don’t like forcing Cursor to follow my design ideas step-by-step. It’s like micromanaging your team — awkward to manage, and it kills productivity. The difference is, the AI team won’t actually be bothered or have its ‘morale’ affected by such micromanaging, meanwhile, a real engineering team will.
Using a Smarter/More Thoughtful LLM Saves You From Round-trip
I encountered this situation a few times. The AI agent was stuck on a problem after numerous LLM requests. In the end, my free quota (of LLM requests) ran out after 2 days.
When tackling complex problems, it’s crucial to opt for a ‘smarter’ LLM. Think of top-tier options such as OpenAI’s o3 models, Google’s Gemini 2.5, or Anthropic’s Claude 3.7. (These models are “smart thinkers” at the time of this essay)
Of course, smarter LLMs are often slow and expensive, so you need to use them wisely. My point is, your intervention becomes necessary when you notice a ‘round-trip’ happening — that is, your coder is stuck going back and forth without good results. When that’s the case, try switching to a smarter LLM and ask your question differently to get a better solution.
Stop AI from Running in Circles by Asking It Questions
When you observe an LLM ‘running in circles’ (or stuck in a round-trip), I usually try to ask it: explain the ‘why’ behind what it’s currently doing and its plan to resolve the problem. Often, this helps the LLM self-correct a flawed thought process and pinpoint the root cause.
It’s very much like the cadence with real human beings. When you rephrase your questions or elaborate on your thought process, great ideas can burst out automatically.
So, don’t just let the AI run in circles without any intervention, hoping it’ll figure things out on the next LLM request after 10 attempts or so. You should stop it, ask questions, and actively guide it forward.
The key point is “Interaction”. You do not even need to know the solution before offering your guidance.
The questions that I often ask:
- Are you sure this is the right solution?
- Tell me what you are doing now?
- Is the problem caused by …?
- What is the difference between solutions 1 and 2?
- Please review your previous step again.
In my experience, when your AI coder gets on the right solution, everything is smooth and fast. Otherwise, it bumps into errors at every step.
Always question the solutions, and ask the AI coder to verify its code.
Don’t blindly trust Agentic AI to always give you correct solutions or bug-free code. I’ve definitely been burned by wrong answers more than a few times. Just as with human engineers, bugs and mistakes are simply a fact of life. That’s exactly why we need tests.
The good news, though, is that AI can usually add tests quickly.
Asking your Agentic AI editor to add tests for the code it generates is like enabling a self-correction loop. I’ve seen my AI coder find and fix issues on its own this way, sometimes without needing my help at all.
Avoid test-loop pitfalls
However, a lot of ‘running in circles’ can happen when it comes to tests. Don’t let your AI coder get stuck digging into the wrong fixes for test failures. Remember the drill: switch to smarter models or provide clear guidance.
And critically, don’t let it modify core architecture or undertake major refactoring without your direct awareness and approval. Be prepared to step in and interrupt when needed.
Grow Your Coding Skills, Expand Your Knowledge via Your AI Coders
View your AI coders not just as output machines, but as unique partners to actively grow your coding skills and significantly expand your technical knowledge. The real learning happens when you know how to ask insightful questions and consciously treat these AI systems as valuable, albeit different, kinds of ‘team members.’
For instance, when an AI generates a piece of code, don’t just copy-paste. Dive deeper. Ask why it chose a particular algorithm or data structure. Inquire about alternative solutions and their respective trade-offs. If a concept or syntax it uses is unfamiliar, prompt it for a clear explanation. You might uncover a more efficient programming pattern you hadn’t considered, learn about a new library that’s perfect for your task, or finally grasp a complex idea that was previously elusive.
Can AI coders replace human coders?
Historically, revolutionary inventions that boost productivity don’t just erase jobs, they reshape them, often creating new roles that demand different skills. AI in coding will likely follow this pattern, transforming developers’ tasks rather than making them obsolete. Those who adapt by mastering AI collaboration will continue to thrive in this evolving landscape.
For those coders who are open to learning new skills and embracing change, the AI bloom will bring new opportunities to thrive. AI coding is still in its early stages. Given time for standards to mature, it will undoubtedly bring new skill set requirements for human coders. Before that day comes, let us maintain a learning mindset, actively use these AI coders, and find our own ways to master them for better productivity.
Summary
In the past few weeks, I enjoyed working with my AI coder. I believe it is the right tool that every programmer should embrace. I look forward to the upcoming Agentic AI tools that can be part of my AI team.
My LinkedIn