00:00:00what really separates the developers who will thrive from those who will get replaced.
00:00:04Ever since AI entered the mainstream, it has started automating a lot of things for us,
00:00:08completely transforming our workflows, as you saw in our previous video that
00:00:12Claude has become an orchestrator of agents. Software developers were the first to adopt
00:00:17it heavily because much of their work involved repetitive code writing which often became
00:00:21inefficient. Now AI is a main part of every developer's workflow, and if you still use
00:00:26AI the same way you did six months ago, you are not keeping up. Upon this scenario,
00:00:31Anthropic released an article discussing trends in software development. As we were talking about
00:00:35it within our team, we found that it was something that was too inherent in our workflow and resonated
00:00:40with us. The software development life cycle is changing dramatically. The cycles that used to take
00:00:46weeks or months are now happening in hours because of AI. The traditional life cycle looked like this.
00:00:51Weeks worth of planning and design, implementation and testing, review, and then the cycle repeated.
00:00:56That changed with AI. Now you only express your intent, and the agent produces an implementation.
00:01:01The only parts where humans are still involved are the review and expressing the intent. The
00:01:05rest is handled by AI agents. This changes what engineering means entirely. Software development
00:01:11doesn't mean writing code anymore. It means orchestrating agents that write code, providing
00:01:16strategic direction, and making sure the system works as intended. Even the onboarding has collapsed
00:01:21from weeks to hours. AI can explore the code base and onboard new joiners immediately. And now,
00:01:26as our focus is on directing agents, everyone is becoming a full stack engineer rather than
00:01:30a specialist in a single domain. Engineers can work with just basic knowledge of their stack,
00:01:35and AI fills the gaps while they lack knowledge. This enables tighter feedback loops and faster
00:01:40learning. Weeks of cross-team coordination become a single working session. This matches exactly what
00:01:45linear's CEO predicted in his article that the middle of the software workflow has been replaced
00:01:50by AI. And if you're still spending your time in the middle phase, you're working against this shift.
00:01:55And this brings us back to the same principle we keep talking about. You need to be effective
00:01:59in your planning and express your intent in clear terms. The skill that matters most now is clarity
00:02:05that is describing exactly what you need and making agents deliver the best product. Before we move on,
00:02:10Team AI Labs is attending the Web Summit 2026 being held in Doha, Qatar. If you are attending
00:02:16or are nearby, this is your chance to meet with the team, connect with us, and learn from us.
00:02:20Looking forward to seeing you there. We've evolved from single agents to multi-agent systems. We have
00:02:26already covered this in our previous video that Claude Code has now implemented multi-agent
00:02:31architecture inside their product. Earlier, the structure used to handle development with a single
00:02:36agent having a single context window and managing all the tasks by itself. The problem was a single
00:02:41context window got bloated fast because there was too much information in its working memory,
00:02:46making it lose focus. Now there's an orchestrator agent that acts like a project manager and
00:02:51delegates tasks to specialist agents. Each of the agents has its own context window and then
00:02:56integrates the output to produce the final result. Even though Claude handles agent spawning and
00:03:00management on its own, we create our own agents for specialized tasks. We use these agents because
00:03:06they were tailored according to our preferences using different Claude models based on the task's
00:03:10difficulty and containing instructions to guide the agent. Sub-agents have gotten better because
00:03:15you can now let them run in the background, handling different aspects of the application
00:03:18simultaneously, speeding up the whole process. Long-running agents will become more capable.
00:03:24Agents have gone from building feature by feature to being able to build complete systems on their
00:03:28own. This started emerging by late 2025, ever since models like Opus 4.5 and GPT 5.2 were released.
00:03:35In 2026, agents will be able to work for days at a time with minimal human intervention. Previously,
00:03:41agents were handling small parts of an application. Now they're building and testing entire
00:03:46applications and systems while verifying if the system is working before moving on to the next
00:03:50feature. We have dedicated a video explaining how to make long-running systems more effective,
00:03:55which you can check out on the channel. With the right tools and workflows, agents are able to plan,
00:04:00iterate, and they recover from failure at scale. This changes the economics of development. In big
00:04:04companies, software accumulates years of technical debt that nobody had time to address. Now agents
00:04:10can actively work through the backlog. This also opens up a path for entrepreneurs. Previously,
00:04:15the main struggle was the skill gap in time. People had ideas but lacked resources to build them. With
00:04:20autonomous agents, startups can now build and deploy products in days. We also use long-running
00:04:25agents for our tasks. Our workflow for long-running tasks uses Claude.md containing instructions. We
00:04:31guide Claude to test after each implementation. For a feature to be well completed, Claude needs a way
00:04:36to verify it's working. We test using agents internally, and for visual testing we use Claude
00:04:41Chrome. Once testing is completed from both the agent's perspective and visual verification,
00:04:46we commit the changes to Git with descriptive messages. This matters because agents tend to
00:04:50modify code and tests that we didn't ask them to. Git lets us roll back easily. We always ask Claude
00:04:56to document the decisions that were made, so commits are clean and ready to be shipped. To maximize time,
00:05:01we ask Claude to break down tasks into smaller, independent units and assign agents to work on
00:05:06them simultaneously. If you want this Claude.md and the agents so that you can use them for your
00:05:11own projects, you can find them in AI Labs Pro. For those who don't know, it is our recently
00:05:16launched community where you get ready to use templates, prompts, all the commands, and skills
00:05:21that you can plug directly into your projects for this video and all the previous videos. If you've
00:05:25found value in what we do and want to support the channel, this is the best way to do it. Links in the
00:05:30description. Human oversight is scaling through intelligent collaboration. As agents are getting
00:05:35better, they can review outputs much faster than we can. We cannot review large-scale outputs that
00:05:40the models are producing at the same speed as agents do, so we are relying on them for all kinds
00:05:45of reviews like security vulnerabilities, architecture consistency, and quality issues.
00:05:50Going through a code base that you didn't write is draining. Agents handle that now. Agents are also
00:05:55learning to ask for help. Rather than blindly attempting tasks, they know when human input is
00:06:00required and ask questions as part of a team. Our team has already noticed this pattern in Claude.
00:06:05When we said that the output looked bad, it asked clarifying questions about what didn't meet our
00:06:10expectations and how it could improve. Oversight is shifting from reviewing everything to reviewing
00:06:15what matters. We only need to review the exceptional cases where problems might emerge. This also
00:06:20addresses the question of AI replacing developers. Even though AI capabilities are expanding, the role
00:06:26of humans remains central. The main change is the shift from writing code to reviewing code and
00:06:30guiding AI outputs. One of Anthropics engineers said the best practice for working with AI is to
00:06:36use it when you know what the right answer should look like. The people who know the answer are those
00:06:41with real software engineering experience who have learned programming concepts the hard way. And how
00:06:46do you know the right answer when you know which method to use for which purpose? For example,
00:06:50for testing you need to use specific approaches. We've already shown you how to use the test-driven
00:06:56approach, white box testing and black box testing. We've also covered visual testing using tools like
00:07:01the Claude Chrome extension and Puppeteer MCP. Also, if you are enjoying our content, consider
00:07:06pressing the hype button because it helps us create more content like this and reach out to more people.
00:07:12Agentic coding is expanding to new services and users that we have never seen before. Earlier in
00:07:172025, AI coding was mostly effective for popular frameworks and often struggled with systems that
00:07:23use legacy languages or frameworks that are not commonly used. That's why the best working
00:07:28applications were React-based because these were the frameworks model was primarily trained on.
00:07:32By 2026, Agentic coding will expand into contexts that traditional development tools couldn't reach,
00:07:38including support for legacy languages like COBOL, FORTRAN and other domain-specific languages. This
00:07:44will make maintaining legacy systems easier by eliminating the need to navigate through
00:07:48old documentation. AI has made development accessible for non-developers, opening up
00:07:53opportunities to non-traditional developers in fields like cybersecurity, operations and data
00:07:58science. The release of co-work is already signaling progress in this direction. The barriers separating
00:08:03people who code and people who don't are becoming increasingly invisible as AI progresses.
00:08:08For example, someone on a security team can use AI to understand unfamiliar code in
00:08:13order to find issues. Research teams have been using it to build front-end visualizations for
00:08:18their data and non-technical employees are already using AI in areas unfamiliar to them like networking
00:08:24and data analytics. This is something our team has already been doing. One of our team members wasn't
00:08:28familiar with Golang but was tasked to make a back-end for a chat application. They turned on
00:08:33plan mode and created a whole plan by answering the questions about the app. Claude built the entire
00:08:38server in one shot, working exactly as intended. This eliminated the need for wasting time learning
00:08:43a new language for just one task. Productivity gains will reshape software development economics.
00:08:48We already mentioned how timelines have been compressed because agents handle the difficult
00:08:53parts. Three factors reinforce each other agent capabilities, orchestration improvements and human
00:08:58experience. Together they compress timelines and change what's viable to build. Projects that were
00:09:03once considered too difficult are now viable, allowing products to enter the market more quickly.
00:09:08Agents help teams work with fewer people. Project timelines are shorter, letting us achieve faster
00:09:12returns on investment. Features that used to take much longer can now be built in a smaller time
00:09:17frame. But before we move forwards, let's have a word from our sponsor, Luma AI. If you've messed
00:09:22around with AI video before, you know the frustration. It usually feels like a slot machine.
00:09:26But Luma AI's new model, Ray3Modify actually changes the game by giving us the modify capabilities
00:09:33that developers have been waiting for. Instead of just prompting and praying, you can now take a
00:09:37video and completely restyle it, swapping environments or lighting while keeping the
00:09:42original motion and physics fully locked. It respects your input data. It's not just generating
00:09:47random noise. It's video to video that maintains structural integrity. Plus with character reference,
00:09:52you can finally keep your subject consistent across different shots, which is usually impossible.
00:09:57It's the first time an AI video feels like a controllable tool rather than just a toy.
00:10:01Make small productions feel huge. Scan the QR code on the screen or check the link in the pinned
00:10:07comment and try Ray3 in Dream Machine today. There is an increase in the number of non-tech use cases
00:10:12across organizations. Teams in sales, marketing, legal and operations can now use AI coding to
00:10:18automate workflows and build tools without any engineering team support. AI agents can operate
00:10:24directly by their guidance and develop systems. People with domain expertise and deep understanding
00:10:29of the problems they face use agents to initiate solutions themselves. For example, someone working
00:10:34in accounting or other departments understands the problems they face better than anyone else.
00:10:39They can instruct agents and have a working solution without waiting for dev team. Our team has already
00:10:44been using Claude in our workflow. We automated the boring non-development work like documentation,
00:10:49ideation and research by using Claude code, letting us focus on the interesting and creative part of
00:10:55our work. Agenda coding improves security defenses and offensive uses. Security and AI are a double
00:11:00edged sword. The same AI that can navigate your code base and help with onboarding is also capable
00:11:06of exploiting its vulnerabilities. Security knowledge is not limited to security engineers.
00:11:10Any engineer can act as a security reviewer, handling hardening and monitoring of systems.
00:11:15Since security engineers are domain specialists, they still need to be consulted. But combining
00:11:20AI with their knowledge, it becomes easier to build, harden and secure applications. While
00:11:25security engineers can defend the applications, there will be offensive use cases too. Last year,
00:11:30we saw a coordinated attack carried out using Claude code and its tools. This means agentic
00:11:35capabilities will evolve the types of attacks we see, making them more intelligent and harmful
00:11:39than ever. Securing systems is going to become increasingly crucial and engineers will need to
00:11:44focus on security from the start. AI agents will play a growing role in cyber defense systems,
00:11:49enabling responses that match the speed of offensive attacks. We need to prepare before
00:11:53attacks happen. We also expect a rise in zero-day attacks, making proactive preparation even more
00:11:58important. When our team creates an app, we use specialized agents for security. These agents
00:12:03handle code review, testing and server-side security, the layer where we control access.
00:12:08Securing applications can be done using different combinations depending on the application,
00:12:12whether it's built-in skills, reusable commands for build purposes or tools from external MCPs.
00:12:18It is better to use an external tool like CodeRabbit because they're built to catch known
00:12:22vulnerability patterns early. That brings us to the end of this video. If you'd like to support
00:12:26the channel and help us keep making videos like this, you can do so by joining AI Labs Pro.
00:12:31As always, thank you for watching and I'll see you in the next one.