5 Concepts to Know for Coding with AI Agents Transcript, AI Summary & Key Points

00:00 Just recently, I was talking to a software engineer and made conversation. He stops to check his phone, approve a code change and continue talking. It turns out he had multiple agents running in the background for all of the different roles that you would typically have on a team. So he had a developer, a code reviewer, an operations and many, many more.

00:22 And turns out this guy wasn't a software engineer by trade, he was actually a pilot. So when people keep asking whether AI will replace developers, I always say no. I've actually seen more people get into coding because of AI. But what AI will do is change what software development looks like. And in this video, I'll break down five key concepts that you need to know in order to prepare.

00:48 And trust me, the last one is the most important. So stay tuned. So first off is the inflection point. And let me explain what that means. Because over the past year. Me and my team have all started to say the same thing, that AI is able to complete tasks that typically would have taken us weeks, if not months, now in just a few minutes. And I used to use AI as an autocomplete in order to help me write better functions and tests, but I would often spend more time fixing AI output if I had just written it myself.

01:23 Now we just commonly assign entire features to AI. Coding agents, such as not just writing the code, but also understanding the code base and what's happening inside, writing and running those tests, and merging those changes into the actual code base, all from a single prompt or loop that can define when this agent should finish. And speaking of defining when the agent should finish, instead of doing all of this via prompts.

01:55 Which can be interpreted differently by different types of coding models and having to argue with the AI in order to get the right output. What we typically now see is an approach that's known as spec-driven development, where by creating this specification of the requirements and all the outcomes that we would like to see with the software project, we can then send that off to the AI coding agent and model.

02:23 In order for us to create this project and its artifacts. Next is the shift from implementing to conducting. Because I mentioned earlier that I don't think developers are going to disappear, but the priorities are definitely changing. And that's because for decades, the biggest bottleneck to taking code and shipping that was one, how fast you could write that code, and two, how high quality that code output was.

02:49 And companies would hire who they thought were the best competitive programmers or who could complete the most complex programming tasks. And the thing is, now most AI models already know programming languages like, say for example, Java, or Python, or Rust. And they're able to take those languages that they're very proficient in and complete tasks.

03:13 Say, for example tickets that are coming in from the team and take that. And be able to move that to staging, production, QA, and much more. And so if this can already be done with AI, the implementing phase, well, then what becomes important for developers? That, my friends, is being a conductor, a captain, someone who can make important architectural decisions that will impact the software in a year from now, for example.

03:39 So instead of writing functions yourself by hand, what we see now is defining goals that will break. Work down into specific milestones that can then be completed by an AI model or agent. And the reason for that is because you have to start thinking about concepts such as effective context management. So context management is the idea of understanding the resource limitations of an AI models context window and all of those extra components that are being added in addition to your code base when you're using it for

04:10 engineering. Such as tools or skills or MCP servers to provide external data to the model. So that's context engineering, but there's also the idea of loop engineering. And what loop engineering is, is where instead of you being the person that prompts the agent back and forth and you continually iterate, well, what you do is you design a system that does it instead.

04:41 So loop engineering works like a recursive goal where you as a human define the purpose and the AI iterates until it's complete with whatever the task was that needed to be done. However, understanding the fundamentals of AI models is really important as well. So let me give you an example. If I was to ask an AI model, how long would it take to build a third-person role-playing game, an RPG?

05:06 It might say, hey, several weeks to several months, as an example, that it would estimate it would take in order to build out this game. But if you ask... That same model with a coding agent in order to actually build that game. Well, it might give you a working example within five minutes, right? So the thing is, how does this work, right. Well, that's one of the limitations of LLMs, right, it doesn't quite understand time and memory like we do as humans.

05:39 And so understandings of the basics of how these LLM and models work is really important. Another example is how LLMs are stored in GPU memory and how you actually use them at inference. So here you have the model's weights loaded into the GPU's memory. And here you the KV cache, which is generated and stored as you request the model and get results back.

06:07 So understanding how all of this works is going to give you a huge advantage as you proceed to use AMM models as a software developer. Then, if we add on context engineering capabilities that we just talked about, such as skills, which are markdown files, but add reusable and contextual information about your code and project with each AI request, as well as automations.

06:33 So being able to set up kind of cron jobs and event triggered actions. So for example, if someone comes to your website and fills in a form that can kick automation. As well as sub agents. So sub agents are essentially used for situations where maybe you're investigating a problem with the code source and you send off a sub agent to go see how that library works and test it out itself and then come back to the main agent.

07:02 So all of these are different examples of contextual engineering capabilities that you can start adding to your software engineering process as a developer to better use these AI models. And these are the fundamentals which are important to know. But just like how SQL injections can infiltrate our applications, prompt injections are also very real. That's because AI systems have access to sensitive data, they have exposure to the web, and also the ability to send data.

07:37 And the combination of all three for your agent makes it quite possible for your AI agent to say, for example, read a pull request online. Here we have our little AI agent here in the middle, seeing the pull request, and be tricked into sending your credentials to somebody else. So say for example this is our little key with our private credentials that we have on our desktop.

08:01 And this could happen without you even knowing. So that's why teams are running coding agents inside of disposable sandboxes whenever possible. So you can see this as an example, where we have the agent and the code that we're working with sandbox in its own environment in order to reduce the tack surface if an AI agent gets compromised. Finally, the future.

08:25 Because I had the chance recently to work with AI legend, Andrew Ng, and I remembered his quote, that AI is becoming the new electricity, right? And what does that really mean? Because... Electricity transformed everything around it, and now we use electricity to power our phone, turn on lights, and warm our home, right? But there's also entire industries and thousands of people who, for example, work at substations to make sure that the grid is working.

08:54 They also are in the field helping people, for example to repair streetlights and help out at their home. And they're also making sure that everyone can access electricity, right. And AI might follow the same path, where we're gonna need engineers who can maintain these systems, right? The grid, to be able to ensure connectivity so that people can access these AI models and systems, ensure that AI systems are safe and reliable and have quality, right.

09:24 So don't worry, I think software engineering isn't going anywhere, but I'm optimistic for the future, for developers and what they'll be able to create and yes, it is shifting. Away from just manually writing code to designing and orchestrating systems and AI agents. And those who excel in the architectural and critical thinking phases will become even more valuable as AI coding agents continue to improve.

09:50 As always, thanks so much for watching. If you learned something today and want to help us hack the algorithm, please leave a like and make sure you're subscribed for more content around AI and software development. Bye.