The year 2025 has felt less like a linear progression and more like a real-time experiment unfolding across the tech industry. We’ve witnessed the capabilities of AI put to the test against human technologists, sparking debates and shifting paradigms at an unprecedented pace. While the year may have begun with AI looking poised to revolutionize software engineering through sheer speed and scale, a crucial evolution has taken hold: the move from what was charmingly dubbed “vibe coding” to the more rigorous discipline of “context engineering.” This isn’t just a change in terminology; it’s a profound recalibration that underscores the enduring, indeed critical, role of human developers.
This evolving landscape is vividly captured in the latest volume of the Thoughtworks Technology Radar, a report that offers a peek into the technologies and techniques shaping our client projects. What stands out is the emergence of new tools and approaches specifically designed to help teams grapple with the complexities of managing context when working with large language models (LLMs) and AI agents. It’s a clear signal: after years of chasing bigger models and faster outputs, the industry is waking up to the understanding that effective context handling is the true game-changer in software engineering and, arguably, in AI itself.
From Vibe to Reality: The Unraveling of Imprecision
Cast your mind back to February 2025, when Andrej Karpathy introduced the concept of “vibe coding.” It was a phrase that resonated instantly, taking the industry by storm. Here at Thoughtworks, it certainly sparked a lively debate; many of us, I’ll admit, approached it with a healthy dose of skepticism. Our concerns were voiced on an April episode of our technology podcast, where we cautiously explored how this seemingly intuitive, yet inherently imprecise, approach might evolve.
Unsurprisingly, given the very nature of something based on “vibes,” antipatterns began to proliferate. The latest Technology Radar, for instance, once again highlighted a worrying complacency with AI-generated code. It’s easy to get swept up in the novelty and speed, but trusting AI outputs blindly can lead to unforeseen issues down the line. Moreover, those early forays into vibe coding inadvertently exposed a degree of complacency about what AI models could truly handle. Users, understandably, began to demand more nuanced, complex outputs, and prompts grew exponentially. Yet, as demands escalated, the reliability of these models began to falter, revealing a gaping hole where comprehensive understanding should have been.
This inherent imprecision and the subsequent reliability issues served as a stark, collective lesson. It became clear that simply nudging an AI in the “right direction” wasn’t enough for robust, production-ready software. This realization was a primary catalyst, propelling us toward a more disciplined and engineered approach to interacting with AI.
Engineering Context: The Cornerstone of Reliable AI Software
The challenges unearthed by vibe coding naturally led to an increasing interest in engineering context. We’ve long been aware of its importance, particularly when working with advanced coding assistants like Claude Code and Augment Code. Providing the necessary context – often referred to as “knowledge priming” – isn’t just a nicety; it’s absolutely crucial. It’s what ensures AI outputs are more consistent, more reliable, and ultimately, lead to better software that requires less rework, reducing those frustrating rewrites and significantly boosting overall productivity.
We’ve seen firsthand the powerful results when generative AI is effectively prepared with appropriate context, especially for understanding complex legacy codebases. It’s a painstaking task for humans, but with the right contextual foundation, AI can make sense of vast, intricate systems. Intriguingly, we’ve even found success in scenarios where full source code access isn’t available, using carefully curated context to glean insights that would otherwise be lost.
Context Isn’t Always About More
It’s vital to remember that “context” isn’t always synonymous with “more data” or “more detail.” This was one of the more counterintuitive, yet profound, lessons we’ve learned from applying generative AI to forward engineering. In this scenario, we’ve observed that AI can be remarkably more effective when it’s abstracted further from the underlying system – that is, when it’s somewhat removed from the granular specifics of legacy code. This is because a higher level of abstraction broadens the solution space, allowing us to better leverage the truly generative and creative capabilities of the AI models we employ. It’s about providing the *right* context at the *right* level, rather than just drowning the model in information.
The Agentic Era Demands Context
The backdrop to many of these recent shifts is the undeniable growth of agents and agentic systems. These intelligent entities are emerging both as products organizations aspire to build and as powerful technologies they want to leverage within their own operations. This burgeoning “agentic era” has squarely forced the industry to reckon with context in a serious way, pushing us decisively beyond any purely vibes-based approaches.
Agents, far from simply executing pre-programmed tasks autonomously, require significant human intervention to ensure they are adequately equipped to respond to complex and dynamic contexts. Their effectiveness hinges on their ability to understand and operate within intricate environments, and without robust context, they quickly become brittle or unreliable. This isn’t just a technical challenge; it’s a fundamental shift in how we design and deploy AI-driven solutions.
To tackle this, a number of context-related technologies are gaining traction, including innovative frameworks like agents.md, Context7, and Mem0. But beyond specific tools, it’s also a question of strategic approach. For example, we’ve found considerable success in “anchoring” coding agents to a reference application. This essentially provides agents with a contextual ground truth, a stable and well-understood benchmark against which they can operate and learn. We’re also experimenting with using teams of coding agents. While this might sound like it introduces additional complexity, it often has the opposite effect, removing some of the burden of having to imbue a single agent with all the dense, multifaceted layers of context it needs to perform its job successfully.
Toward Consensus and Collaborative Ground Truths
As this space matures, we can expect practices and standards to embed, bringing much-needed structure to the chaos of innovation. It would be remiss not to highlight the significance of the Model Context Protocol, which has rapidly emerged as a leading standard for connecting LLMs and agentic AI to diverse sources of context. Similarly, the agent2agent (A2A) protocol is blazing a trail in standardizing how different agents interact with one another, paving the way for more sophisticated multi-agent systems.
Whether these specific standards ultimately win out remains to be seen, but their emergence signals a collective desire for interoperability and reliability. Crucially, beyond the protocols and tools, we must also consider the day-to-day practices that enable us, as software engineers and technologists, to collaborate effectively. After all, AI needs context, but so do we. Simple yet powerful techniques, like establishing curated shared instructions for software teams, might not sound like the hottest innovation, but they can be remarkably effective for fostering cohesive human collaboration, especially when dealing with highly complex and dynamic systems.
There’s also a fascinating conversation to be had about what these changes mean for agile software development. “Spec-driven development” is one idea gaining traction, suggesting a more upfront, context-rich specification process. However, the core agile tenets of adaptability and flexibility remain paramount. The challenge lies in building robust contextual foundations and ground truths for AI systems while simultaneously preserving our ability to iterate, pivot, and respond to change.
Software Engineers: The Architects of Context
Without a doubt, 2025 has been a monumental year in the evolution of software engineering as a practice. There’s a lot the industry needs to monitor closely, but it’s also an incredibly exciting time to be involved. While lingering fears about AI-driven job automation may persist, the very fact that the conversation has decisively shifted from questions of sheer speed and scale to the intricate challenges of context places software engineers squarely at the heart of this revolution.
It will, once again, be down to human ingenuity, expertise, and collaborative spirit to experiment, to learn, and to adapt. The future of intelligent software development, and indeed the effective integration of AI into our lives, hinges on our ability to master context. The software engineer isn’t just a coder; they are the architect of understanding, the builder of contextual intelligence. The future, truly, depends on it.
