How to Apply Agentic Coding to Solve Problems

Coding with agents has become the single most effective approach for me to solve problems. Most problems I encounter at work can be solved effectively by utilizing agents. This is in contrast to manually solving tasks or coding up a solution yourself.

In this article, I’ll give a high-level overview of how I approach problems and solve them using Claude Code. As an engineer, you’re essentially a problem solver. Your job should be something like:

1. Discover and identify the most important problems to solve
2. Come up with a solution to the problem
3. Execute

And this doesn’t just apply to programming tasks. It also applies to tasks in marketing, sales, and customer management. I know this because this is what I do every single day at work as a Data Science Lead of a series-A funded startup.

Why you should solve problems with Claude Code

I always think it’s important to know why you do something. If you take any engineering major at university, this is a mindset they’ll instill in you.

Always understand the why of what you’re doing

You should solve problems with Claude Code simply because it’s usually the most effective way to solve a problem. Now, of course, you can use Claude Code alternatives, such as Cursor, though I’ll refer to Claude Code throughout this article, because it’s the tool I use.

However, Claude Code isn’t only effective in step 3 of the problem-solving process (executing on the solution). It’s also effective in discovering and identifying problems, along with how important they are. It’s also super effective at coming up with solutions to the problems you’ve discovered.

You can, for example, discover important problems by:

• Having Claude analyze production logs and inform you of any issues
• Give Claude access to your CRM system to clean up and notify you of the most important areas to focus on
• Give Claude access to all your social media posts and their performance, to analyze what works and what doesn’t

After discovering issues like the ones listed below, it’s important to prioritize them. One of the simplest tools for this is the value effort graph.

You simply list all problems to solve on a 2D graph, with the axes being the value you get from solving the problem and the effort required to solve the problem. After doing this, you simply pick the problems that have the highest value, relative to the effort required to solve them.

How to Apply Claude Code to discover and solve problems

In this section, I’ll go through the three steps of the problem-solving process, which I highlighted earlier: identify and prioritize problems, come up with a solution to the problem, and execute on the solution.

To keep the article organized, I’ll have one subsection per step of the problem-solving process.

Step 1: Discover and prioritize problems

Earlier in the article, I highlighted some specific ways to discover important problems with Claude Code.
I think the most important factor in finding important problems to solve is:

Give Claude Code access to all the information you have access to

If you don’t give your coding agents this access, you simply can’t expect them to perform well. You can imagine if you had to solve a production bug without being allowed to look at the logs. Of course, you wouldn’t be able to solve it, because you cannot understand what the problem is, without looking at the logs.

Thus, make sure your coding agent has access to all relevant resources:

• Project management tools like Linear
• Notes from Notion
• Log groups in AWS
• GitHub to look at commits
• Browser access to reproduce issues
  And probably many more platforms. The point is: if you use the platform to solve a problem, you should give your coding agents access as well.

Step 2: Coming up with solutions to problems

Okay, at this point, you’ve done the hardest part: You’ve identified an important problem you need to solve. In software engineering, we often say that:

Finding the bug, why it happens, and reproducing it is the hardest part. Solving it from there is easy

So if you’ve gotten here, you should be excited that you only have the simple part yet. Of course, how you come up with a solution depends a lot on the problem you’re trying to solve.

If it’s a production bug, you can often just prompt Claude Code with:

Problem X is happening in location Y. I believe it’s because of Z. Look at the CloudWatch log groups †o understand why the problem is happening, and come up with a plan to solve it

This is enough for over 50% of the problems I work with, and it works well to solve production bugs and implement simple features. If the problem is a bit more complicated, you often need to iterate several times on the plan Claude Code creates for you, read thoroughly through it, and make sure everything looks correct. This all ties back to general techniques and approaches I’ve previously discussed to get the most out of Claude Code.

---

For other problems, you might need to prompt Claude Code differently, or make sure it has access to the correct tools. If you’re, for example, analyzing your CRM tool, you’ll need to provide API access to the tool and provide Claude Code with all the access it needs.

In general, I urge you to be liberal with the access you provide your coding agent, because without proper access, it simply cannot solve problems for you. This is exactly the same as if you were to ask humans to solve a problem. If they don’t have access to the correct platforms and tools, you cannot expect them to be good problem solvers.

Still, however, you should of course make sure the agent cannot perform any destructive actions. If you’re giving it AWS access, for example, you should probably log in with Viewer access only, instead of an admin user. In general, you can simply follow good security practices.

Step 3: Executing on solutions

The last part of the problem-solving process is to execute the solution. If you’ve come up with a good plan with Claude Code, you can simply tell it to execute on its plan.

Again, this works for a majority of the problems I apply Claude Code to. It’s gotten so good that it one-shots the solution.

However, I still encounter more complex problems where this doesn’t work. Sometimes you need to iterate several times with Claude Code. For example, when I create marketing material such as:

• Scripts for webinars
• PDF carousels for LinkedIn
• Posters

I often have the agent make an initial design and start reviewing and iterating on it. In general, this is a great technique when working on complex problems. Come up with a decent initial solution, test it, review what works and what doesn’t, and iterate. By simply repeating this a few times, you’ll usually end up with a great solution.

Conclusion

In this article, I’ve discussed how I apply Claude Code to solve problems. In general, my mindset is that whenever I encounter a new problem, I think: “How can I apply Claude Code to solve this problem for me. A lot of times, you’ll realize that you can solve a problem with a simple prompt. Other times, it requires more iterations, either to discover the right problems, prioritize the problems, create a plan to solve the problem, or to execute on the problem. In general, I strongly urge you to have the mindset of iteration in place, where you start off with a decent solution, and iterate on it until you’re happy with it. Most problems don’t require a perfect solution.

👉 My free eBook and Webinar:

🚀 10x Your Engineering with LLMs (Free 3-Day Email Course)

📚 Get my free Vision Language Models ebook

💻 My webinar on Vision Language Models

👉 Find me on socials:

💌 Substack

🔗 LinkedIn

🐦 X / Twitter