have revolutionized the best way I program. After I first discovered coding again in 2019, I wrote all of the code, character for character. In hindsight, I’m grateful for this expertise, as a result of problem-solving mindset coding taught me.
Nonetheless, with AI brokers, I imagine I’m at the least 10x more practical as an engineer than I was. It is because I’m using AI brokers to carry out as lots of the repetitive, mundane duties as doable that I used to must do. Beforehand, I needed to:
- Manually discover the indentation error in Python
- Carry out lengthy analysis by Googling for solutions
- Carry out massive refactors all manually
And an extended checklist of different duties I don’t spend a lot time doing now. Some would say you shouldn’t let AI do all of the give you the results you want, as a result of it makes you a worse coder. I’d counter that AI is just doing the mundane repetitive work, whereas I can carry out probably the most cognitive-straining duties, organizing and orchestrating the AI brokers.
On this article, I’ll offer you an perception into three of crucial strategies I implement to successfully make the most of AI brokers for coding. I imagine these strategies each make me much more environment friendly as an engineer, and it additionally maximizes the potential of AI brokers.
I’m at all times on the lookout for methods to be a more practical engineer, so if in case you have extra options, I’d tremendously recognize any suggestions!
All through the article, I’ll check with each Cursor and Claude Code. I’m not sponsored by any of them, and it’s merely the instruments I exploit for my agentic coding.
Why it is best to use AI brokers for coding
I first need to begin off with why it is best to use AI brokers when coding. The principle argument is that you are able to do extra in much less time.
AI brokers can help you do extra, in much less time
Options that used to take 5 hours to implement can now be realistically carried out and examined in quarter-hour.
Bugs that took you an hour to determine and half-hour to unravel can now merely be solved by your agent in 5 minutes. With the Linear MCP, you may even simply copy the difficulty URL and have your agent learn the request, discover the error, implement an answer, and create a ready-made PR for you.
That is merely superb. You’ll be able to learn studies on how AI solely will increase effectiveness in 5% of implementations. Nonetheless, in the event you begin utilizing agentic coding instruments and the strategies I’m presenting on this article, I’ll assure you see a noticeable change in your productiveness.
I’m pushing code greater than ever earlier than, and not using a vital improve in bugs. That’s the literal definition of what makes a programmer more practical.
Method 1: Cursor Planning Mode
The primary and possibly most essential method I make the most of is the planning mode every time implementing a brand new function. Planning mode is offered in the entire well-known agentic coding instruments, reminiscent of Claude Code and Cursor. In Cursor, you may choose it by urgent Shift-Tab whereas prompting your agent.
Planning mode makes it so the agent makes a plan for an implementation, both for a brand new function, a bug report, or no matter you need to carry out in your codebase. This plan is made by the agent studying by way of your request and the code repository as a way to decide the most effective method to unravel your request.
Plan mode then writes the plan to a Markdown file, and would possibly ask you some questions on tips on how to implement your request:
- Which language would you like the implementation in? Python or TypeScript?
- Ought to the implementation be backwards suitable?
- Monolith structure, or microservices?
Plan mode is so efficient as a result of pure langauge is inherently ambigoutous
That is the rationale we got here up with a programming language: A chunk of code is deterministic, and at all times outputs the identical, given the identical enter. There is no such thing as a ambiguity in a bit of Python code, for instance.
Nonetheless, now we’ve reverted to coding by way of pure language, as a result of the brokers are implementing code as a substitute. Nonetheless, we’d like an method to eradicate ambiguity, the place the plan mode is available in helpful.
Method 2: Cursor Browser Actions
Cursor browser actions is one other very helpful method I’ve began actively using recently. Cursor browser actions permit your Cursor agent to carry out actions whereas working in your implementation. These actions will be:
- Open a URL
- Press a button
- Learn console logs
That is an unimaginable device to assist the agent resolve one-shot issues, all by itself, as a substitute of you having to manually take a look at implementations within the browser and replica console logs containing errors.
As a substitute, you may merely ask Cursor to open the URL, click on round, and ensure there are not any points with the implementation. You primarily make Cursor carry out an end-to-end take a look at for you, which is tremendous efficient at discovering challenges in your implementation.
It saves a variety of time to immediate Cursor to open the browser, click on round, and examine the browser logs each time I add a brand new implementation.
Method 3: Test logs with Claude Code
One other helpful method I make the most of closely is to examine logs with Claude Code or Cursor.
I exploit Claude Code to examine logs in the event that they’re unrelated to an implementation I’m engaged on. For instance, if a doc is instantly caught within the processing pipeline, with out me having modified something within the related code just lately.
I exploit Cursor sometimes to examine logs every time I’m engaged on an implementation. I can, for instance, immediate Cursor to ship a take a look at occasion to a Lambda perform, and examine the CloudWatch logs to ensure all the pieces was processed as anticipated.
I beforehand spent a variety of time within the AWS console, manually navigating to the related log group and looking out the log group, which takes a very long time. Thus, I began prompting my coding brokers to examine the logs for me as a substitute, which saves me 10-60 minutes per day: an unimaginable effectivity achieve.
Having brokers examine logs is tremendous helpful, and has virtually no draw back. Checking logs is normally a easy job that you simply simply must do. It’s not intellectually difficult or one thing you need to spend time on. Thus, using brokers to scan by way of logs is an excellent useful use case of coding brokers.
Checking logs is a straightforward and repetitive job: an ideal job to have coding brokers carry out, whilst you do extra useful work
Should you use brokers to scour by way of your logs, it’s essential to supply your agent with as a lot background data:
- What are your log teams known as
- What are the desk names
- What are the S3 bucket names and prefixes
This protects you a variety of money and time, as a result of your agent doesn’t must checklist your entire infrastructure as code and discover the related service to look by way of. I’ve talked about this idea in my article on effective usage of AGENTS.md.
Conclusion
On this article, I’ve mentioned three of the principle strategies I exploit to successfully make the most of coding brokers. I imagine that the usage of coding brokers and these strategies has made me at the least 10x more practical as an engineer, from an general perspective. It’s revolutionized the best way I work, and saves me unimaginable quantities of time in my day-to-day work. I imagine being efficient at using AI instruments can be extremely essential for the programmers of the longer term.
👉 My Free Assets
🚀 10x Your Engineering with LLMs (Free 3-Day Email Course)
📚 Get my free Vision Language Models ebook
💻 My webinar on Vision Language Models
👉 Discover me on socials:
🧑💻 Get in touch
✍️ Medium
