Framework (For your brain) for Improving Agentic Software Engineering Processes
Week #29 of grinding on my #Object #Storage service, and I’m super stoked to share some insights on leveling up agentic software engineering processes. This week, I teamed up with GitHub Copilot to co-author a feature, and it got merged today (PR). Sounds like AI just does your job for you, but trust me, it’s not that easy. Verifying the code is correct is a real pain, especially when you’re not totally clear on your own implementation details. Knowing those details in the back of your head lets you make smarter calls on performance, abstractions, and all the small stuff that matters.
Get Tejus Chandrashekar’s stories in your inbox
Join Medium for free to get updates from this writer.
Before I let Copilot take over, I put in some serious prep, around 3-4 hours over the week. Here’s what I did:
- Locked down the requirements: I spent a good chunk of time figuring out exactly what I wanted. Check it out here: https://github.com/vitality-ai/warpdrive/issues/78.
- Hand-coded integration tests: I carefully hand coded all the integration tests for my APIs to make sure they were rock-solid. See them here: https://github.com/vitality-ai/warpdrive/pull/84.
- Experimented with TLA+ for context: Got inspired by Kiro’s blog on the future of software development and messed around with TLA+ to give my coding agent some extra context. Peek at it here: https://github.com/vitality-ai/warpdrive/tree/main/tla. No hard proof it helped, but it felt like a solid move. Kiro’s framework vibes are legit, and my small project worked like a charm.
What Worked?
- Clear requirements are a game-changer: Laying out the details like I did in the GitHub issue made all the difference. Usually, if you’re vague, Copilot just spits out random code, and you’re stuck cleaning up the mess. This time, it nailed it and the code in https://github.com/vitality-ai/warpdrive/pull/86 is exactly how I’d have written it myself. Total win!
- Tests are your best friend: Hand-coding those integration tests gave me confidence that Copilot’s work wasn’t just fluff. Everything passed my test suite, so I could relax knowing it was solid.
- TLA+? Still figuring it out: I’m not fully sold or maybe I did not understand yet, Marc’s take about LLMs getting TLA+ . It was fun to try, but I’d love some solid metrics to back it up. Hoping to see that soon. For me the question here is, the integration of predicate logic and Markov chains, which are inherently deterministic, with the stochastic nature of LLMs and how do these structured frameworks align with the randomness characteristic of large language models.
This is just the start of my framework for agentic software engineering. It’s a small project, but it’s already showing some serious promise. Stay tuned for Part 2-unless I get sidetracked again, haha!
Note: Some formatting was handled with the help of an LLM, but the writing, thoughts, and work are all mine. That said, I’ll admit these days it’s tough to tell who did what.
PS: Grok seems to generate nice images.
