00:00:00We are in a new era of software development. Developers are shipping products at a speed
00:00:04we have never seen before. However, a problem has emerged. Traditional workflows do not hold
00:00:08up when agents are involved. This raises an important question. What does the developer
00:00:13role look like now? A recent article by the CEO of linear caught my attention. Linear is a project
00:00:18management tool that helps teams organize and track their work specifically for modern software
00:00:23development. These insights come from someone who has lived through the transition from traditional
00:00:27workflows to the AI driven systems of today. This article made me rethink more than just the tools
00:00:33we use. It made me rethink how we build products entirely. We've got a lot to talk about today
00:00:37because this information fundamentally changes how we build with AI. The middle of software work is
00:00:43disappearing and the center of software is actually moving. To understand what the middle is, let us
00:00:47look at how work was divided before AI development. It started with the beginning phase. This included
00:00:52all of the requirements gathering and planning phases. In this phase, we created plans for what
00:00:57we were going to build. Then came the middle. This was where we converted the plan to the actual
00:01:01product and it was the part where writing the code was involved. This was the part that took the most
00:01:05time of all. It took weeks, months or even a year to deliver a quality fully working setup. This was
00:01:11also the part where details got mixed up the most because of translating intent or conveying ideas
00:01:16from one person to another. After the code was written, the end part included various forms of
00:01:20testing and reviewing against the original requirements. The middle was the part that contained
00:01:25the most friction but the CEO says that is no longer going to be the case. It is because the
00:01:30middle work, the implementation and the coding part is actually being replaced by AI. Now we don't have
00:01:35to touch code ourselves at all. This is because coding agents have become so powerful that they
00:01:40are able to produce code from context and task planning alone. It is now more about using the
00:01:45agents the right way and supervising their work than writing code. If you've been watching our
00:01:50videos regularly, we have taught and demonstrated many different ways you can use coding workflows
00:01:55to produce production level apps. You can do this by just supervising the agents without having to
00:01:59code a single line yourself. IDEs have become more of a code viewer than a writing tool. This
00:02:04change is really apparent to me because as a developer, my go-to tool for writing code has now
00:02:09become a tool for reviewing the code the agent produces. Now I just go to VS code to review or
00:02:14add comments so the AI agent can implement the commented features. I very rarely have to change
00:02:19anything or write code myself now because agents are highly capable. But this only works if the
00:02:23agents are able to understand the intent. Therefore our work as developers has essentially shifted from
00:02:28writing code to supervising it. You've probably noticed we build a lot in these videos. All the
00:02:33prompts, the templates, the stuff you'd normally have to pause and copy from the screen. We've put
00:02:38it all in one place. We've recently launched AI Labs Pro where you get access to everything from
00:02:43this video and every video before it. If you found value in what we do and want to support the channel,
00:02:48this is the best way to do it. Links in the description. Since AI has taken over most of the
00:02:53coding work, this leads to a question. What is left for us? The answer is focusing on the new craft of
00:02:59refining the intentions of what to build. The way you can do that is by treating planning as your
00:03:03primary job. You need to clearly understand the problem you are trying to solve. You need to know
00:03:07what your customer actually wants and how people will use your app. This has become even more
00:03:12important now. You are no longer relying on humans who can interpret intentions from poor planning.
00:03:17Instead, you are relying on AI agents that blindly implement whatever you instruct them to do. Whether
00:03:23you are building a mobile app or a web app, you need to know exactly what you want to build. Without
00:03:27that clarity, you cannot do meaningful planning with the agent's planning modes. Planning is vital.
00:03:32As we have emphasized in our previous videos, only good plans lead to good implementation. It does not
00:03:38matter which agent you are using. Planning is very important because it controls the outcome of the
00:03:42agent. Take as long as you need. Keep refining the plan until it fully satisfies your needs and meets
00:03:47your expectations. This will ensure that your app turns out the way you want. Until three months ago,
00:03:52we never relied on bypass permission mode for building because agents used to hallucinate
00:03:56despite a good plan. Now the agents are so reliable that after refining the plan, I just turned the
00:04:02bypass permission mode on and let the agent implement the specs in a single run. We also
00:04:06saw that even the creator of Claude Code starts his implementations with plan mode. If the plan
00:04:12is good enough, you can let the agents build the app in one shot without worrying about messy
00:04:16implementations. I also spend a significant amount of time making sure that what I am building is
00:04:21fully documented. I do not cram it all into a single document so the agent can navigate through
00:04:26the plans easily. I use different documents for each category such as risk assessments, mitigation
00:04:31and text specs. I list constraints and trade-offs in a separate document. This is how the agent
00:04:35understands what is acceptable in terms of performance, cost and time. This approach leads to
00:04:40much more controlled development. After all the requirements have been verified, the next step is to
00:04:45actually manage the agent and get what we want. But before we talk about that, here's a quick word from
00:04:50our sponsor. Dart AI. Managing complex software projects often involves more administrative overhead
00:04:56than actual coding. Dart is not just a standard project management tool. It is an AI native
00:05:00workspace designed to automate busy work for developers. With the context-aware AI chat, you can
00:05:05even create tasks and edit documents just by talking naturally. Beyond AI chat, you can even onboard
00:05:11agents like cursor to execute work. Dart gives them the context to actually write your code. The real
00:05:16power lies in its AI guidelines feature. You can configure global rules like instructing the AI to
00:05:22always format technical specs with specific goals and requirements headers and Dart enforces this
00:05:27structure across every chat, task and document it generates. For us, the AI skills feature is a game
00:05:33changer. You can define custom commands like a generate project skill that automatically creates
00:05:38a populated task list, assigns priorities, estimates sizing and drafts a project brief in seconds.
00:05:44Start automating your project management today by checking out Dart AI at the link in the pinned
00:05:49comment. You are no longer just a coder. Your work is more centered around supervising agents than
00:05:54actually writing code. Writing code has become less about constructing a solution and more about
00:05:58setting up the conditions for a good solution to emerge. So how do you create the right environment
00:06:03for agents to produce quality outcomes? The answer is context engineering. The next big skill you
00:06:08need to learn is not a specific web development stack like MERN or MEAN. Instead, it is context
00:06:14management. We've consistently seen that without proper context management, it does implement the
00:06:18features we prompt it to, but it doesn't follow any constraints or rules it had to match the
00:06:22implementation to. We need to ensure the context is managed properly. When the agent is given the
00:06:27right information with minimal noise, it understands the task more clearly. It produces better
00:06:32implementations and delivers exactly what you want. Managing the context involves using a set of
00:06:37components like reusable commands, skills, markdown files, MCPs and sub-agents. There is no single
00:06:43right way to do it. You should use multiple methods that work well for what you are trying to build.
00:06:47You need to create a workflow that suits your project. We have dedicated an entire video
00:06:52demonstrating how you can build workflows with context management. This ensures the model you
00:06:56are using gets the right context and can produce high quality applications. If you want to follow
00:07:01along, all the resources for that video are available in AI Labs Pro. An agent's work is
00:07:06only as good as the context-driven environment it operates in. The more it is connected directly
00:07:11to customer feedback and supported by a structured workflow, the better it can perform. We need to
00:07:16create such an environment because it does not happen automatically. For this reason,
00:07:20Claude has connectivity with Slack so that teams can directly report errors. This creates valuable
00:07:25feedback loops, which even the creator of Claude code himself used. Large teams are already producing
00:07:30high quality AI generated code. The creator of Claude code claimed that in the past month,
00:07:35100% of his contributions were effectively written by Claude code itself. This does not happen just
00:07:41by giving it a prompt. It requires a set of workflows and orchestrated patterns to make
00:07:46it possible. Even the CEO of Microsoft admits that AI now generates 20% to 30% of Microsoft's
00:07:52integrated code across all languages. There is especially notable progress in Python and C++.
00:07:58Structure in tools works the same way for both humans and agents. It reduces uncertainty by
00:08:03clearly defining what is expected and what capabilities exist. If you are using AI agents
00:08:08without structure, you are only using a fraction of their potential. The structure can take many
00:08:13forms. This includes a Claude.md file for overall project guidance and a change log to track changes.
00:08:19You can also use reusable/commands or specialized skill.md files with scripts and references.
00:08:25Additionally, you can use plugins and MCP tools to extend the agent's capabilities.
00:08:29But knowing these tools is not enough. The right combination matters. Every project requires a
00:08:34different setup, so you have to build one based on your project's needs. With the right balance,
00:08:39you will get results just the way you want. Our job is not done after planning and delegating
00:08:44tasks to agents. Now, as I mentioned that I let Claude code work on dangerously skip permission
00:08:49mode, it does save a lot of time, but it requires our time and attention towards something else.
00:08:53The pressure shifts towards the end of the cycle. Reviewing the code becomes more important.
00:08:58Code that is not reviewed can lead to degraded performance and high costs. You can use structured
00:09:02workflows to make reviewing easier. This will lead to fewer bugs and save you from issues later on.
00:09:07Now, testing is not just going to your agent and saying test my app for all the issues. It involves
00:09:12several approaches to improve the process. One method is test-driven development. We ask the
00:09:17agent to write test cases for the feature we want to implement without writing any code initially.
00:09:22Once the tests are written, I clear the context and start a new window. This ensures the agent
00:09:26loses context on how it wrote the tests. I ask Claude to run the tests and they fail because
00:09:31no code has been written yet. Now that I know the tests are working correctly, I ask Claude to
00:09:36implement the route. I ensure that it does not modify the tests. This way, the agent has a clear
00:09:41goal to iterate toward. In TDD, tests are written before the code, but testing should also happen
00:09:46after the code is written. For that purpose, there are many forms of testing. I use blackbox testing
00:09:51and create user stories. These act as detailed guides on how users will actually interact with
00:09:56the system and how those interactions might trigger errors. Blackbox testing evaluates the
00:10:00functionality of an application based on requirements without looking at the code itself.
00:10:05I then use the Claude Chrome extension to perform the testing and ask it to iterate over each user
00:10:10story, section by section. Blackbox testing mainly identifies functionality issues. For performance
00:10:16testing, we also need whitebox testing. This is where we actually look at the code, not just the
00:10:21output. We trace how the code is implemented and reason about its architecture. For whitebox
00:10:25testing, I used an XML document containing multiple sections and subsections of tests. This document
00:10:31acts as a guide for Claude on how to navigate through written code and how to find architectural
00:10:36issues. To simplify my testing, I used a custom command that executes the tests in the document
00:10:41which I placed in the testing folder. This command lists the instructions for initializing the tests,
00:10:46how to log the results into a file in a structured format, and at the end, how to generate a final
00:10:51report. This slash command made whitebox testing easy for me because it contains the structured
00:10:56prompt for testing. Since the middle is disappearing and the focus is shifting more toward the beginning
00:11:01and the end, we need to rethink our priorities. What we need to prioritize now is forming the
00:11:05right intent through planning and requirement assessment. We must also ensure that the outcome
00:11:10meets expectations through thorough testing and review processes. Those developers who master these
00:11:15principles will be the ones leading the future. That brings us to the end of this video. If you'd
00:11:20like to support the channel and help us keep making videos like this, you can do so by using
00:11:24the super thanks button below. As always, thank you for watching and I'll see you in the next one.