Summary
Highlights
The speaker welcomes the audience to an event on a systematic approach to working with AI for developers in 2026. He notes the rapid evolution of technology, particularly AI, in programming over the past year. He mentions that while many are already using AI, some are still resistant or have had negative experiences. The goal of the session is to share knowledge to help developers understand their current position in adopting AI and identify areas for improvement.
The speaker emphasizes the speed at which AI tools are updating. He checks for updates to 'Open CD' before the stream begins and finds a new version. He notes that if an AI agent project hasn't been updated in three days, it feels abandoned, highlighting the exceptionally fast development cycle of AI software compared to other software.
The speaker differentiates between two main types of AI use in programming: 'development of agents' and 'development with the help of agents' (AI-assisted development, authentic programming, or AI coding). He explains that the latter is more common for developers using AI in their daily coding tasks, while the former involves creating the AI agents themselves. He provides an example of an AI assistant at Hexlet that helps users with courses and projects, noting the challenge of teaching AI not to give direct answers.
The speaker clarifies that 'agents' in many development contexts are often more like workflows with fixed processes rather than truly autonomous systems. He describes how AI is used at Hexlet to analyze student questions, replacing a team of methodologists by identifying common problems and recommending course improvements. He draws a parallel to the evolution of chatbots, where initial low-level APIs have given way to platforms that enable non-technical users to build complex AI systems.
The speaker addresses the exaggerated claims about AI writing 80% or more of code. He refers to an analysis of GitHub commits that found only about 3% of commits are explicitly identified as AI-generated, though he acknowledges the true number is likely higher due to unflagged AI usage. He highlights that official statistics are often based on non-representative surveys and can be misleading, but the trend of AI adoption is clearly accelerating.
The speaker dismisses concerns about AI growth plateauing, comparing it to Moore's Law, where technological limitations are overcome by new approaches. He addresses the niche for local AI, noting it's mainly for large businesses with strict security requirements that build their own AI models in-house. He clarifies that general developers won't need to configure these local models themselves; platform teams will provide them as integrated tools.
The speaker discusses confidentiality and proprietary code with AI. He states that for his own business, code confidentiality is less concerning than the value of the product itself. He highlights that large companies have significant concerns with licensing and data security, especially regarding personal data, emphasizing the need for strict rules and avoiding direct AI access to production databases.
The speaker outlines the progression of AI in development: from basic IDEs and Stack Overflow to current AI agents and future autonomy. He envisions AI systems working in parallel, across devices, and integrating with other development tools. He notes that while multi-agent systems and sub-agents sound promising, they can increase costs and introduce complexity due to error propagation. He encourages developers to continuously use AI for research and planning, even on the go, to optimize their workflows.
The speaker discusses various ways to interact with AI agents. He distinguishes between standalone AI editors like Cursor and Antigravity, and AI plugins for existing editors like VS Code. He personally finds standalone editors less practical, believing plugins are more adaptable and will ultimately prevail. He cautions against the hype around new AI products, suggesting many will either be acquired by larger companies or fade away.
The speaker compares the effectiveness of AI agents embedded in editors versus command-line interface (CLI) tools. He states that conceptually, all agents are similar and quickly adopt each other's best features. He advises developers to stick with tools integrated into their preferred development environment (e.g., JetBrains or VS Code) unless there's a clear advantage to using a separate CLI.
The speaker advises caution against the hype surrounding AI, particularly claims from individuals who haven't coded in years. He suggests looking at the actual impact of AI-generated products, noting that many have low adoption rates without proper marketing. He mentions that even well-known developers sometimes hide their use of AI due to perceived stigma. He explains that AI models from major companies like Google and Antropic often provide their own CLI tools, while independent tools like Open Code offer universal access to various models.
The speaker identifies three key skills for working effectively with AI agents: 1) Task formulation: goes beyond simple prompt engineering to include clear thinking and problem understanding; 2) Learning and patterns: understanding how different AI models behave and when to use each one, adapting to their evolving capabilities; 3) Integration: effectively connecting AI agents with existing tools and systems to provide context and automate complex tasks. He emphasizes the importance of having a clear end goal before starting to code with AI.
The speaker elaborates on advanced integrations, segmenting them into tools, MCP (Messaging, Communication, and Performance), and skills. Tools refer to integrating with external command-line utilities for stateless operations, which are often more efficient than stateful MCPs. Skills combine these tools to create complex functionalities. He notes the growing trend of libraries and frameworks providing AI-ready documentation and skills for better integration, making it easier for AI to understand and utilize them.
The speaker discusses the learning curve for developers using AI. He suggests that while AI can assist in many areas, developers should still practice coding manually to maintain their fundamental understanding and critical thinking. He compares it to learning to play an instrument – tools can assist, but fundamental skill is essential. He also highlights that AI agents' rapid development makes it necessary to continuously adapt and learn their capabilities, often through trial and error.
The speaker shares an anecdote about using an AI agent to refactor authentication logic. He initially gave a broad prompt, which led to incorrect changes due to the AI's limited understanding of the project's specifics. By resetting the context and providing a new prompt focusing on incremental changes without altering interfaces, he achieved a better outcome. This illustrates the importance of managing AI context and understanding its limitations, especially for complex tasks.
The speaker begins a live demonstration using Open Code to check website performance. He showcases the variety of models available and their integration capabilities, including connecting to existing systems like GitHub Copilot. He explains the concept of 'planning mode' where the AI strategizes without making code changes, contrasting it with the 'build mode' which executes actions. He also highlights the importance of managing context to avoid AI's 'hallucinations' or irrelevant actions.
The AI agent in planning mode analyzes the website for performance and accessibility issues using Chrome DevTools. It identifies a problem with a button's accessibility due to incorrect HTML structure. The speaker then switches to a more powerful model (GPT) to generate a plan for fixing the issue. He emphasizes that while a basic model might struggle, a sophisticated one can quickly diagnose and propose accurate solutions, illustrating the significant difference in AI capabilities.
The demonstration continues with the AI executing the fix. The speaker observes that the cheaper 'Minim Max' model struggles to implement the solution correctly, attempting to modify outdated YAML files instead of the current JSON files. This highlights the importance of using the right model for the right task and that even with a good plan, a less capable AI might fail in execution. He then switches to ChatGPT to correctly implement the fix, showing how a more advanced model efficiently applies the solution.
The speaker summarizes key takeaways from his AI demonstration. He emphasizes that AI often fixes symptoms rather than root causes, so developers need to guide it to address fundamental issues. He also highlights the vastly different performance between basic and advanced AI models, stressing that using a weak model can lead to frustration and a negative perception of AI. He notes that competent human oversight is still crucial for effective AI utilization.
The speaker shares his experience using AI as part of job interviews. He found that AI does not help unqualified candidates; instead, it often exposes their lack of understanding more clearly. He uses real-world coding tasks from his projects to assess how candidates use AI for problem-solving. This approach helps him gauge a candidate's critical thinking, problem-solving skills, and ability to effectively interact with AI tools.
The speaker announces a new month-long course on 'AI for Developers' starting on the 30th. The course will cover an introduction to AI agent development, practical skills for using tools and integrations, effective work with agents, and building autonomous systems integrated with platforms like GitHub. A key component of the course is a project where participants build an application from scratch using their chosen technologies, with AI handling 90-95% of the coding, emphasizing a 'design first' approach and API-driven development.
The speaker answers various audience questions. He discusses the use of Serene MCP for optimizing token consumption, noting historical trends indicate that CLI utilities often outperform stateful MCPs in cost efficiency. He also addresses code review challenges with AI-generated code, indicating a need for more standardized approaches. He clarifies that fundamental programming skills remain crucial, as AI enhances rather than replaces them. He also touches on how AI influences job requirements, forcing developers to adapt and become more versatile to stay competitive.