How AI Agents have changed Developer Workflow

There is no denying that AI agents have changed the way we do things as developers.

1. Understanding a New Codebase

Before: We would review the documentation (if available; it has always been very helpful) and the README, then follow the instructions to set up the project on our machine. We would then review the architecture diagram to understand the high-level architecture, and we would go to the specific parts of the codebase where each component is implemented to understand how each part is implemented and to get a feel for the coding standards for that project. We might even trace the flow of data using pen and paper to better understand a specific part where we are performing our tasks.

Now: Although the above methods are still helpful for us, we might use an AI agent to summarize the docs, generate an architecture diagram where it does not exist, explain specific components, trace out, and display data flow on a particular part. We can even have a conversation with the AI agent, asking it questions and describing our assumptions to better understand the codebase in a conversational style. Of course, if the codebase is very large, it takes better planning and context feeding to get the desired outcome from the agent.

2. Debugging

Before: When we receive a bug report, we try to reproduce it on our machine. If we can't, it is even harder to debug since there might be deadlocks or race condition issues that present themselves in a production environment, while being invisible to us. So we might resort to utilizing our observability metrics and using stress testing to figure out the problem. We then try to trace the flow of data through the code using print statements or a debugger and identify where it went wrong. We would then plan and implement a fix based on our understanding of the issue.

Now: We can utilize an AI agent in every step of finding out where the issue lies, figuring out what went wrong, and planning and implementing a fix. We do need to control how much we are outsourcing to the AI agent so that we can be sure about the result and we do not lose context of what we are doing.

3. Writing tests & docs

Before: We would need to identify every case we need to test for and write every line of test code for each case. We also needed to manually write a detailed documentation, which could actually result in stale and unupdated documentation due to the time and effort it takes to produce and update the documentation.

Now: This is by far the best utility of AI agents to save our time on writing verbose tests and detailed documentation. We can actually ask the AI agent to write tests, and we can read the test and look for uncovered edge cases, which we can iterate with the AI agent to fill out. We can also let the AI agent write docs, and we can read, iterate, and verify with improved efficiency compared to what we would have done before AI agents. 

4. Implementing new features

Before: We would plan the way the feature would work, and we would draw out the new components we needed to add, the existing components we need to extend, and how they would work together. We would also try to structure everything in a way that is easy to read and test. After that, we would write the code for the new feature by hand.

Now: When we plan a new feature, we can involve an AI agent to poke holes in our plans, remind us of our oversight, and even give us new alternatives to explore and decide on. We can then feed the AI agent our finalized plan and let it write the code for us. We then need to review the code and validate that everything is according to plan. This improves our efficiency a lot by focusing our attention on the plan and architecture instead of typing out every single line of code. This works better if we provide the agent with clear instructions and an unambiguous plan to the best of our abilities.

We still need our judgment and technical skills to produce quality work. We can also utilize AI to accelerate our progress while doing our work.