Dev lifecycle improvments with AI
What can LLM agents help with in your coding lifecycle
As LLMs have improved, there is a lot of hype as to their impact on software development. Over the year IDE integrated tools like Github Copilot, Cursor have improved a lot as the underlying models improved. In addition to that there are completely standalone coding agents like Claude Code, GPT Codex and Google Jules/Antigravity have emerged with more capabilities. I haven’t worked a lot with the standalone coding agents, but have used Github copilot extensively and have seen the improvment.
So here are the concrete things that LLMs are helpful in improving the dev lifecycle from my experience. I’ve mostly used Claude Sonnet 4.5 and Opus 4.5 in my evaluation
Faster understanding of legacy code
LLMs can help us understand the workflows and technical organization of legacy code which speeds up understanding of the code. Initially they were poor in understanding large code bases because of context limitations. But once the agents learnt patterns to search, they execute multiple searches and calls and understand the code and can document the architecture for us to understand the flow. This helps us get up to speed with old code faster and also to become productive with new features or fixes in the code.
Peer to think through
LLMs take the rubber duck debugging concept much farther. You have much capable duck to talk through your issue, get concrete suggestions and alternative points to think through to debug/plan/re-architect your project.
Faster code construction
This one is obvious and the first usecase of the copilots. Now in the ‘agent’ mode given a requirement then can plan and one-shot the code required for implementing a feature. They can also compile the code and run them to check if the code is running. They can also execute tests to see if if the requirements are satisfied.
Tests
This brings us to the next big advantage. We can iteratively come up with the testcases that need to be created and the agent can write the tests and run them instantly increasing the confidence in the code not just that time but even for the future.
Faster Migration
Another avenue where LLMs shine is in the case where you have to migrate your code from one version to another like java 11 to java 17 or spring boot 2.x to 3.x or AWS SDK 1.x to 2.x etc, then these agents can help you migrate the code super fast. Just ask them to create a plan and then keep working through the plan as required. I needed to open some of the files and ask them to be converted and it did a great job and I was able to migrate a huge application in 1 day. This is a big win.
Error analysis
This is a corollary to the duck debugging, if you give an exception stack trace, the agent can identify the cause based on the exception and suggest some possible reasons and mitigations for the same.
Custom operational tools
With the faster speed of development, it is now possible to develop custom tools to help with manual operational stuff without having to spend a lot of time. It is my 20% project to develop new tools to improve our operational posture.
Pitfalls
Is everything rosy then? - Here are some of the things that don’t work as well or might be counterproductive
It might have subtle bugs that are not immediately evident, so you have to review the code carefully and test all scenarios before merging the changes.
You might forget syntax and rely too much on the LLM. Not an easy solution here but it might be ok?
It can sometimes just comment out the code causing errors or failing tests after trying to fix them a few times, so it is important to keep looking at what the agents are doing.
It might not stick to the existing architecture and want to make it similar to generic architecture it is trained on aka complex solutions. So it is imporant to make sure it complies with the architecture of the existing app.
If it codes out a whole new app, then you might not really understand what all it is doing. You can ask it to document the architecture with diagrams to understand it.
Even with all these pitfalls, there is no doubt that it has double the productivity of software engineers if not more. It does make it more fun to work with code.
Having good unit/integration tests and blue/green deployments will help the teams take advantage of it. Marin Kleppman predicts that formal verification will go mainstream with AI programming. I think that is a great outcome though there is a learning curve.