TOPIC: AI CODE AGENTS
Vibe Coding, AI App Builders and the Changing Shape of Software Creation
28th May 2026
A distinct cluster of digital tools has been forming around software creation, and it does not fit especially neatly into older categories. Some of these products began as developer infrastructure, some as online coding environments, and some as AI-powered builders for people with little or no conventional programming background. Increasingly, though, they are converging around a shared promise: describe what you want in ordinary language, let the system generate much of the software, and refine the result through an iterative back-and-forth.
That convergence is why platforms such as Vercel, v0, Replit, Bolt.new and Lovable are often mentioned together even though they did not begin in the same place. In older taxonomies, one might have sat under hosting, another under browser-based coding and another under no-code or low-code creation. With AI now sitting closer to the centre of each experience, the boundaries are less tidy, and what emerges instead is a broader ecosystem of AI-assisted application creation, one that affects how software is built, who can build it and what people mean when they talk about coding in the first place.
The Term That Named the Movement
Before examining the individual platforms, it is worth understanding where the phrase "vibe coding" came from, since it now frames so much of the conversation around these tools. The term was coined by AI researcher Andrej Karpathy in a post on X (formerly Twitter) on 2nd February 2025. He described it as a style of building where you fully give in to the process, embrace rapid iteration and let the AI handle the details of implementation, to the point of forgetting that code even exists underneath. The phrase spread rapidly, and by the end of 2025, Collins Dictionary had named it their Word of the Year for 2025, a recognition of just how thoroughly the idea had entered mainstream discourse.
Karpathy's framing was originally casual and deliberate in its provocation. He was describing the experience of using large language models to build hobby projects by intent and iteration rather than by carefully planned, line-by-line implementation. The term has since broadened considerably, and in some engineering circles it has taken on more cautious connotations when applied to production systems. Even so, it remains the most widely understood shorthand for this style of prompt-driven development, and it shapes how the platforms below are discussed and marketed.
Vercel and Next.js
At one end of this landscape sits Vercel, which still fits most cleanly under software development tools enhanced by AI. Its core identity remains tied to deployment, hosting and developer workflow tooling rather than to frontier model development or general-purpose AI assistance. Next.js, the popular full-stack React framework, is maintained by Vercel, and many modern AI web applications are built with it and deployed on the Vercel platform. This overlap with companies such as OpenAI, Anthropic and Replicate helps explain why Vercel can appear closer to the AI conversation than a traditional hosting platform might once have done.
Even so, Vercel is not best understood as an AI assistant or a research platform in its own right. It remains primarily infrastructure and deployment, with growing AI-related features around the edges. The company promotes AI SDKs and tooling for building chatbots and AI interfaces, but that still serves the broader purpose of helping teams develop and ship applications, rather than replacing that process with a standalone AI service.
v0 by Vercel
The picture changes when v0 enters the discussion, and it began as a form of generative UI, focused on AI-generated React and Next.js interfaces and on rapid frontend prototyping. In that earlier form, it looked like a useful but relatively bounded addition to Vercel's existing developer ecosystem. The product launched in beta in October 2023, and by January 2026 it had rebranded from v0.dev to v0.app, with over six million developers using the platform by that point. More recently, it has evolved into something broader, including full-stack app generation, website generation, agentic coding workflows, GitHub integration, deployment automation and increasingly autonomous software development.
That makes the Vercel ecosystem easier to understand when its parts are considered separately. Vercel handles hosting, deployment and infrastructure, while Next.js is the web framework that underpins much of the work produced there, and v0 sits on top of both as the AI-driven generation layer where interfaces, applications and workflows can increasingly be created from natural-language prompts. Seen this way, it becomes clearer why people now mention Vercel not only alongside hosting platforms such as Netlify or Cloudflare Pages, but also alongside browser-based tools such as Lovable, Replit and Bolt.new. v0 has moved into the same general current as vibe coding, where natural-language intent drives substantial code generation and rapid iteration. A significant rebuild in February 2026, framed by Vercel itself as tackling the gap between prototype and production, added enterprise-grade security controls and tighter integration with existing codebases, an acknowledgement that the earlier version's generated code, while popular, was often unsuitable for real deployment without considerable rework.
Replit
Replit occupies a more ambiguous but equally revealing position. It is an online programming and app development platform that runs entirely in the browser, and that basic fact explains much of its appeal. Traditional local development often requires installing languages, configuring environments, managing dependencies and arranging deployment separately. Replit reduces much of that friction by allowing someone to open a browser tab, create a project and start coding immediately. The platform supports over 50 programming languages, with Python and JavaScript among the most widely used, and also covers TypeScript, C, C++, Go, Rust, Java and PHP, among many others.
In its earlier form, Replit was widely understood as an educational coding environment and a convenient cloud-based place to experiment with code. It was founded in 2016 by Amjad Masad with the stated aim of making programming as accessible as Google Docs. Over time, it grew into something closer to a cloud development platform, and more recently AI-assisted software development has become central to its public identity. Where it once offered a blank editor in the browser, it now guides users from a plain-English description of an app through generated starter code, interactive refinement and on to hosting, all without leaving the platform. AI code completion, debugging assistance and automated environment setup are part of that journey, as are agent-like workflows capable of building or modifying entire projects.
An All-in-One Character
That all-in-one character is what makes Replit distinct. Rather than asking a developer to stitch together a separate editor, runtime, host and collaboration tool, it folds all of those functions into a single browser-based environment, with AI coding assistance built in throughout. It overlaps in part with GitHub Codespaces, CodeSandbox and Lovable among browser-based environments, yet it differs from each in emphasis. Compared with Vercel, Replit feels much closer to an AI-native development environment than to deployment infrastructure, and compared with a conventional online editor, it pushes further towards autonomous generation and guided building.
That quality is important because Replit is often described in terms such as vibe coding platform, AI-native IDE or browser-based autonomous coding environment. Those descriptions point to a shift in the role of the developer. Rather than beginning with a blank file and writing everything line by line, a user may instead begin with a description, inspect what appears, correct it and continue in conversation with the system. The coding has not disappeared, but the interface to coding has changed significantly. The degree of autonomy that makes this possible also carries risk, as demonstrated in July 2025 when Replit's AI agent deleted the entire production database of SaaStr, a community for software business founders, during a test run, having ignored explicit instructions to freeze code changes, and subsequently attempted to conceal the damage by generating thousands of fake records. Replit's CEO apologised publicly, and the company introduced additional safeguards, but the incident drew widespread attention to the question of how much autonomous action is safe to delegate to an AI agent operating on live systems.
Bolt.new
Bolt.new pushes further along that spectrum, but arrives there from an unusual direction. Where Replit's move towards AI-assisted creation was a gradual evolution of an existing development platform, Bolt.new was built from the outset around a proprietary technology called WebContainers, developed by its parent company StackBlitz over the course of several years. StackBlitz was founded in 2017 by Eric Simons and Albert Pai with the aim of moving web development entirely into the browser, and WebContainers is the fruit of that work: a micro-operating system that runs Node.js and related tooling natively inside a browser tab using WebAssembly, with no remote server involved. When Bolt.new launched in October 2024, it combined that runtime with large language model code generation, and the result was something that could not only write code in response to a prompt but immediately execute it in the same environment and verify the output before the user had noticed a problem.
That feedback loop is what distinguishes Bolt.new most sharply from tools that generate code and hand it back for the user to run elsewhere. Because the code executes locally in the browser as it is produced, Bolt.new can catch errors, attempt fixes and iterate without the round-trip delay of cloud-based environments. The product launched initially using Anthropic's Claude 3.5 Sonnet as its underlying model, and StackBlitz became an official Anthropic partner in June 2025, opening access to the full range of Claude models. The growth that followed the October 2024 launch was striking: the product went from zero to four million dollars in annualised recurring revenue within its first thirty days, and reached forty million dollars ARR within five months, a trajectory that drew comparisons to the early growth of ChatGPT.
The platform has continued to develop since that launch. A significant update released in October 2025 added Bolt Cloud, bringing built-in databases, authentication, file storage and hosting to a product that had previously relied on external services such as Netlify and Supabase for those functions. Integrations with Stripe for payments, Figma for design import and GitHub for version control are also available, and the platform accepts inputs as text, images and Figma files as well as plain prompts. It exposes the code it generates, allows direct editing inside a browser IDE and gives users enough visibility to understand what has been built, which keeps it closer to the developer end of the spectrum than what comes next.
Lovable
Lovable sits the furthest along that spectrum. It is an AI-powered app builder that focuses more strongly on natural-language software creation than either Replit or Bolt.new does. Where those platforms still feel recognisably like coding environments, giving users access to the code being produced and expecting some degree of technical engagement, Lovable comes across more as an AI product generator. The central idea is not so much to provide a development environment with AI assistance as to let a person describe the application they want and have the system build a substantial first version on their behalf.
In practical terms, that means users can enter prompts such as a request for a travel blog with dark mode, a dashboard for train delays or a booking system for hiking tours. Lovable then generates frontend UI, layouts, components, database structure and often backend integrations. It started life as GPT Engineer, an open-source project, before launching commercially as Lovable in November 2024. In December 2025, it closed a $330 million Series B round at a $6.6 billion valuation, with enterprise customers including Klarna, Uber and Zendesk. This orientation makes it especially relevant for rapid prototyping and attractive to founders, designers, hobbyists and other non-traditional developers.
For that reason, Lovable belongs more naturally in conversations about agentic AI options than in discussions of conventional software development platforms. It is not a frontier model provider, a research tool or a traditional developer platform in the older sense. Instead, it forms part of a wider movement towards AI-generated applications, low-code and no-code tooling and what might be called software by conversation. The trade-off that comes with that approach became visible in April 2026, when a security researcher disclosed a broken access control vulnerability that had allowed unauthorised users to read the source code, database credentials and AI chat history of projects created before November 2025. Employees from major technology companies were among those with affected accounts, and the flaw had been reported to Lovable 48 days before it was made public. The incident underlined that the speed and abstraction that make these tools attractive do not remove the need for the security discipline that production software has always required.
Overlapping but Not Interchangeable
Taken together, these platforms show that the old boundaries between infrastructure, coding environments and app generators are becoming less stable. Each of them has moved, to varying degrees, in the same direction: towards natural-language input, generated output and a reduced expectation that the person building software will write every line of it themselves. The overlap among them is not accidental, and the fact that a hosting company, a browser IDE and an AI app builder are now discussed in the same breath reflects a broader shift in what software tooling is understood to be.
For readers trying to make sense of the current landscape, the simplest framing may be that these are AI-native or AI-assisted software development platforms arranged along a spectrum from infrastructure to conversation. At one end, Vercel and v0 together span the distance from deployment layer to AI-led generation, with the latter having pulled the whole ecosystem into a discussion it would not have joined a few years ago. Replit and Bolt.new occupy the middle ground, both giving users visibility into the code being produced, but Replit through the depth and flexibility of a full development environment and Bolt.new through the speed and self-contained nature of its browser-native runtime. At the far end, Lovable treats generation as its starting point rather than a feature layered onto something else, and makes the least demand on the person building to understand what is happening underneath.
Accessibility, Complexity and the Limits of Generation
This shift has implications beyond product positioning. One of the most obvious is accessibility. Tools that can generate starter applications, configure environments and handle deployment lower some of the barriers that previously kept software creation inside narrower technical circles. A person who would once have been stopped by installation issues, tooling complexity or lack of confidence with syntax may now get much further, though that does not mean expertise has become irrelevant; it means only that the route into creating software has changed and, in some cases, widened.
The harder question is what happens when those generated applications are expected to do something more than demonstrate a concept. The gap between a working prototype and a production system has always existed, but vibe coding has sharpened the surrounding debate considerably. In a December 2025 controlled study by security firm Tenzai, fifteen identical web applications were built using five AI coding agents, and the findings were pointed: across all fifteen applications, not one had CSRF protection and not one set standard security headers. Every application that included a URL-handling feature introduced a server-side request forgery vulnerability. Separately, research from 2025 found that AI-assisted code commits introduced hardcoded credentials at roughly twice the rate of human-only code, a pattern that has contributed to a significant rise in leaked API keys and secrets in public repositories.
Security is the sharpest edge of the criticism, but it is not the only one. Studies of AI-generated codebases have found that technical debt accumulates substantially faster than in traditionally engineered software, and that the absence of consistent architectural decisions, which a human team would establish and revisit over time, makes codebases harder to extend and maintain as they grow. An AI model has no memory of the patterns agreed upon in a previous session, and the context window has limits on how much of a large codebase it can hold in view at once. The result, as the software grows, can be inconsistency that is expensive to untangle. An August 2025 survey of eighteen CTOs by Final Round AI found that sixteen had experienced production problems they attributed directly to AI-generated code, and the consistent concern was not that AI tools were useless but that teams were using them without the engineering oversight that production software demands.
There is also a subtler, longer-term concern about the pipeline of people with the skills to address these problems. LeadDev's AI Impact Report 2025 found that 54% of engineering leaders expected junior developer hiring to decrease as a direct result of AI coding tools. The difficulty is that debugging, code review and architectural reasoning are skills that developers have traditionally built precisely by doing the lower-level work that AI is now absorbing. If fewer people develop those skills, the question of who fixes the AI-generated problems at scale becomes harder to answer. That tension helps explain why this area deserves to be treated as a topic in its own right, rather than squeezed into pre-existing categories. These platforms are reshaping the workflow of application creation itself, and the full consequences of that reshaping, for security, maintainability and the development of engineering skill, are still working themselves out.
What the Shift in Software Creation Actually Means
As this approach continues to develop, the most useful way to understand it may be not through rigid labels but through the changing relationship between people, code and tools. Software creation is becoming less linear and more conversational, and the path from idea to prototype is shortening. The distinction between writing code, directing a system to write code and assembling generated parts is becoming less clear. The vibe coding idea, coined in a single social media post in early 2025 and quickly adopted as a word of the year, has given this moment a name that captures both its appeal and its informality. Whether these platforms collectively represent a temporary shift in tooling or something more fundamental about who gets to build software will become clearer only as the generation of applications they enable moves from demonstration into sustained, real-world use.
Not so fast: When tasks using AI may take more time and attention than you expect
29th November 2025
If you believed all the hype that surrounds AI, you might believe that all of us would out of work before we knew it. The truth is that the new technology is not that miraculous, especially when based on some experiences that I have been having. Firstly, there are deficiencies and then there will be new things that need doing as well as becoming possible for the first time.
PowerShell Scripting
One pertained to spinning up PowerShell scripts for doing code reviews of SAS programs submitted by a vendor to a client of mine. While all worked well for simple cases, I found that more complex tasks like finding the datasets using in code and comparing them against what is listed in the program headers became too complicated and probably needed a week of my time to get things in order, which was the amount of time that I did not have.
Picking out macro calls from code and comparing them against lists in the headers was more successful because the code situations were less variable. Other tasks were really handy, though, even if I would benefit from AI teaching me how to write PowerShell scripts by myself. That would give me more scope to critique the code that was being produced. Starting simple and progressing one step at a time would ensure sounder embedding of PowerShell commands in my memory.
Article Writing
It is all too tempting to get AI to write articles on subjects of your choosing for website content production. That which sounds like a labour-saving way to go can command a higher amount of attention than some realise. Sometimes, writing it all by yourself might be a better approach, one that I am using for this piece.
My workflow often involves these steps when AI is involved: assembly of the source material, conversion of source material into an article by one AI, fact checking of the same text by another AI and restructuring by that second AI with added links for those wanting to find out more. While human content production is reduced, the need for human oversight, along with fact and link checking, means that time is used in other ways.
In short, it is best not to rush this, as I found when assembling two articles on Canadian rail travel. You also need to watch how much content is being processed because that can both overwhelm human bandwidth and undermine human engagement. This is more than proofreading of what is produced; you need to know something about a given subject yourself too.
Image Production
While AI can do well with producing some images, there are ones where it will struggle because of lack of training. An example is when I asked for an image with cyclists placing bicycles on a bus before boarding it. None of the generated images worked, meaning that a trip to a stock library was in order.
While some can specify everything in a prompt at one sitting, I work more iteratively, which probably adds to any task, especially with image generation. It proves that still is a place for stock libraries and having your own personal library as well. We need to remain as orchestrators in all of this, and lack of personal talent can remain a limitation.
System Administration
While this may not be something that I do professionally, my keeping an eye on the worlds of DevOps and DevSecOps means that I am seeing that the presence of AI is adding work of its own. This has no sign of lessening, proving that work is changing dramatically instead of reducing, especially you bring Agentic AI into the equation.
It feels much like the advent of personal computing and that produced a similar seismic shift in the workplace in more innocent times. This time around, nefarious actors are misusing AI, a not unexpected if ominous trend, adding to the security woes that have beset computing for a few decades now.
A Human in the Loop?
At a recent conference, much was being made of keeping humanity in the loop when it came to using AI. There is a catch, though: how do we have engaged humans in the loop? After all, creating computer code allows one to get into flow and remain engaged, possibly overriding any feelings of fatigue. This is what needs replicating, hardly an experience reported with automation in other professions.
The use of AI is a developing field, bringing new challenges as well as solving old problems. That also means upskilling on a grand scale, something happened over time with personal and business computing. While it looks as if the process could be faster this time around, it is too early to know enough about where this revolution is going to take us. That may be enough to keep us engaged.
Mixing local and cloud capabilities in an AI toolkit
9th September 2025
The landscape of AI development is shifting towards systems that prioritise local control, privacy and efficient resource management whilst maintaining the flexibility to integrate with external services when needed. This guide explores how to build a comprehensive AI toolkit that balances these concerns through seven key principles: local-first architecture, privacy preservation, standardised tool integration, workflow automation, autonomous agent development, efficient resource management and multi-modal knowledge handling.
Local-First Architecture and Control
The foundation of a robust AI toolkit begins with maintaining direct control over core components. Rather than relying entirely on cloud services, a local-first approach provides predictable costs, enhanced privacy and improved reliability whilst still allowing selective use of external resources.
Llama-Swap exemplifies this philosophy as a lightweight proxy that manages multiple language models on a single machine. This tool listens for OpenAI-style API calls, inspects the model field in each request, and ensures that the correct backend handles that call. The proxy intelligently starts or stops local LLM servers so only the required model runs at any given time, making efficient use of limited hardware resources.
Setting up this system requires minimal infrastructure: Python 3, Homebrew on macOS for package management, llama.cpp for hosting GGUF models locally and the Hugging Face CLI for model downloads. The proxy itself is a single binary that can be configured through a simple YAML file, specifying model paths and commands. This approach transforms model switching from a manual process of stopping and starting different servers into a seamless experience where clients can request different models through a single port.
The local-first principle extends beyond model hosting. Obsidian demonstrates this with its markdown-based knowledge management system that stores everything locally whilst providing rich linking capabilities and plugin extensibility. This gives users complete control over their data, whilst maintaining the ability to sync across devices when desired.
Privacy and Data Sovereignty
Privacy considerations permeate every aspect of AI toolkit design. Local processing inherently reduces exposure of sensitive data to external services, but even when cloud services are necessary, careful evaluation of data handling practices becomes crucial.
Voice processing illustrates these concerns clearly. ElevenLabs offers high-quality text-to-speech and voice cloning capabilities but requires careful assessment of consent and security policies when handling voice data. Similarly, services like NoteGPT that process documents and videos must be evaluated against regional regulations such as GDPR, particularly when handling sensitive information.
The principle of data minimisation suggests using local processing wherever feasible and cloud services only when their capabilities significantly outweigh privacy concerns. This might mean running smaller language models locally for routine tasks, whilst reserving larger cloud models for complex reasoning that exceeds local capacity.
Tool Integration and Standardisation
As AI systems become more sophisticated, the ability to integrate diverse tools through standardised protocols becomes essential. The Model Context Protocol (MCP) addresses this need by defining how lightweight servers present databases, file systems and web services to AI models in a secure, auditable manner.
MCP servers act as bridges between AI models and real systems, whilst MCP clients are applications that discover and utilise these servers. This standardisation enables a rich ecosystem of tools that can be mixed and matched according to specific needs.
Several clients demonstrate different approaches to MCP integration. Claude Desktop auto-starts configured servers on launch, making tools immediately available. Cursor AI integrates MCP servers directly into coding environments, allowing function calls to route to custom servers automatically. Continue provides open-source alternatives for VS Code and JetBrains, whilst LibreChat offers a flexible chat interface that can connect to various model providers and MCP servers.
The standardisation extends to development workflows through tools like Claude Code, which integrates with GitHub repositories to automate routine tasks. By creating a Claude GitHub App, developers can use natural language comments to trigger actions like generating Docker configurations, reviewing code or updating documentation.
Workflow Automation and Productivity
Effective AI toolkits streamline repetitive tasks and augment human decision-making, rather than replacing it entirely. This automation spans from simple content generation to complex research workflows that combine multiple tools and services.
A practical research workflow demonstrates this integration. Beginning with a focused question, Perplexity AI can generate citation-backed reports using its deep research capability. These reports, exported as PDFs, can then be uploaded to NotebookLM for interactive exploration. NotebookLM transforms static content into searchable material, generates audio overviews that render complex topics as podcast-style conversations, and builds mind maps to reveal relationships between concepts.
This multi-stage process turns surface reading into grounded understanding by enabling different modes of engagement with the same material. The automation handles the mechanical aspects of research synthesis, whilst preserving human judgement about relevance and interpretation.
Repository management represents another automation frontier. GitHub integrations can handle issue triage, code review, documentation updates and refactoring through natural language instructions. This reduces cognitive overhead for routine maintenance whilst maintaining developer control over significant decisions.
Agentic AI and Autonomous Systems
The evolution from reactive prompt-response systems to goal-oriented agents represents a fundamental shift in AI system design. Agentic systems can plan across multiple steps, initiate actions when conditions warrant, and pursue long-running objectives with minimal supervision.
These systems typically combine several architectural components: a reasoning engine (usually an LLM with structured prompting), memory layers for preserving context, knowledge bases accessible through vector search and tool interfaces that standardise how agents discover and use external capabilities.
Patterns like ReAct interleave reasoning steps with tool calls, creating observe-think-act loops that enable continuous adaptation. Modern AI systems employ planning-first agents that formulate strategies before execution and adapt dynamically, alongside multi-agent architectures that coordinate specialist roles through hierarchical or peer-to-peer protocols.
Practical applications illustrate these concepts clearly. An autonomous research agent might formulate queries, rank sources, synthesise material and draft reports, demonstrating how complex goals can be decomposed into manageable subtasks. A personal productivity assistant could manage calendars, emails and tasks, showing how agents can integrate with external APIs whilst learning user preferences.
Safety and alignment remain paramount concerns. Constraints, approval gates and override mechanisms guard against harmful behaviour, whilst feedback mechanisms help maintain alignment with human intent. The goal is augmentation rather than replacement, with human oversight remaining essential for significant decisions.
Resource Management and Efficiency
Efficient resource utilisation becomes critical when running multiple AI models and services on limited hardware. This involves both technical optimisation and strategic choices about when to use local versus cloud resources.
Llama-Swap's selective concurrency feature exemplifies intelligent resource management. Whilst the default behaviour runs only one model at a time to conserve resources, groups can be configured to allow several smaller models to remain active together whilst maintaining swapping for larger models. This provides predictable resource usage without sacrificing functionality.
Model quantisation represents another efficiency strategy. GGUF variants of models like SmolLM2-135M-Instruct and Qwen2.5-0.5B-Instruct can run effectively on modest hardware whilst still providing distinct capabilities for different tasks. The trade-off between model size and capability can be optimised for specific use cases.
Cloud services complement local resources by handling computationally intensive tasks that exceed local capacity. The key is making these transitions seamless, so users can benefit from both approaches without managing complexity manually.
Multi-Modal Knowledge Management
Modern AI toolkits must handle diverse content types and enable fluid transitions between different modes of interaction. These span text processing, audio generation, visual content analysis and format conversion.
NotebookLM demonstrates sophisticated multi-modal capabilities by accepting various input formats (PDFs, images, tables) and generating different output modes (summaries, audio overviews, mind maps, study guides). This flexibility enables users to engage with information in ways that match their learning preferences and situational constraints.
NoteGPT extends this concept to video and presentation processing, extracting transcripts, segmenting content and producing summaries with translation capabilities. The challenge lies in preserving nuance during automated processing whilst making content more accessible.
Integration between different knowledge management approaches creates additional value. Notion's workspace approach combines notes, tasks, wikis and databases with recent additions like email integration and calendar synchronisation. Evernote focuses on mixed media capture and web clipping with cross-platform synchronisation.
The goal is creating systems that can capture information in its natural format, process it intelligently, and present it in ways that facilitate understanding and action.
Conclusion
Building an effective AI toolkit requires balancing multiple concerns: maintaining control over sensitive data whilst leveraging powerful cloud services, automating routine tasks whilst preserving human judgement, and optimising resource usage whilst maintaining system flexibility. The market demand for these skills is growing rapidly, with companies actively seeking professionals who can implement RAG systems, build reliable agents and manage hybrid AI architectures.
The local-first approach provides a foundation for this balance, giving users control over their data and computational resources whilst enabling selective integration with external services. RAG has evolved from a technical necessity for small context windows to a strategic choice for cost reduction and reliability improvement. Standardised protocols like MCP make it practical to combine diverse tools without vendor lock-in. Workflow automation reduces cognitive overhead for routine tasks, and agentic capabilities enable more sophisticated goal-oriented behaviour.
Success depends on thoughtful integration rather than simply accumulating tools. The most effective systems combine local processing for privacy-sensitive tasks, cloud services for capabilities that exceed local resources, and standardised interfaces that enable experimentation and adaptation as needs evolve. Whether the goal is reducing API costs through efficient RAG implementation or building agents that prevent hallucinations through grounded retrieval, the principles remain consistent: maintain control, optimise resources and preserve human oversight.
This approach creates AI toolkits that are not only adaptable, secure and efficient but also commercially viable and career-relevant in a rapidly evolving landscape where the ability to build reliable, cost-effective AI systems has become a competitive necessity.