From virtual assistants to autonomous decision-making systems, AI is being reimagined to tackle increasingly complex tasks. At the heart of all these efforts is agentic AI, and behind that is the concept of context engineering.
In a recent episode of the Box AI Explainer Series podcast, host Meena Ganesh and Box Chief Technology Officer Ben Kus discuss this pivotal advancement that’s redefining how AI agents operate. Context engineering is not just a dressed-up version of prompt engineering. It’s a scaffolding of information and tools that enables AI agents to produce the best results and even work in tandem with each other.
In this insightful conversation, Ganesh and Kus explored why context engineering isn’t just the next step — it’s the foundation of building truly effective AI agents.
Key takeaways:
- Context engineering gives AI agents the information and tools they need to produce the best results in even the most complex scenarios
- Prompt engineering and context engineering differ, but they have a relationship — and both are critical to getting the right results
- Context engineering is quickly becoming fundamental to building effective AI-driven products and services in the enterprise
"Context engineering is the discipline by which you assemble information and tools so an AI agent can then perform a complex task," Kus explains. “It’s basically the art and science of programming your agents.”
In the world of AI agents, successful output often requires multiple iterations, sometimes connecting multi-agent systems. AI agents need a lot of context to get the job done properly, and context engineering asks, “What kind of environment, data, and tools do AI agents need to perform seamlessly?”
The goal of context engineering is to equip agents with everything they need to know before tackling a user’s request.
Prompt engineering vs. context engineering
Taking a step back, Kus describes how “in the old days” (in other words, last year — or even last week), “To get an AI model to do what you want, you’d prompt them.” You’d give them examples of what you’re looking for and maybe a few instructions. And then you might spend a bit of time tweaking the results with further model prompts.
But at that point, the expectation was that AI would produce a fairly simple answer to a problem — aka a “single shot.” As AI agents have evolved to handle multi-faceted, iterative tasks, the spotlight has shifted to context engineering.
Unlike traditional approaches that focus on a single interaction between a user and an AI tool, context engineering considers the bigger picture, offering AI agents a detailed framework to operate within.
The value of context engineering
Imagine you're trying to analyze a company’s financial health. As Kus explains, "Even if you’re intelligent — maybe the smartest in this area — you need context to answer a question like that.”
If you task an AI agent to do it for you, it still needs information about the company’s operations, industry benchmarks, recent performance reports, and more. But if the agent becomes overloaded with irrelevant snippets or fails to receive enough meaningful details, it can’t deliver actionable insights.
Information overload and ambiguity are just as confusing to AI agents as they are to humans — possibly even more so. This is why context is critical to AI agents. They need to know what information is important. Context is not just about instructing an agent; it’s about ensuring the agent knows what data to consider and has access to that data.
“One of the ways you know that you have good context engineering or bad context engineering is whether or not you’re getting what you’re looking for from your agent,” Kus says. “If you don’t get what you want, sometimes it’s not because the agent’s not smart. It’s because it doesn’t have the data it needs and hasn’t been engineered well to answer the question.”
The AI agentic ecosystem
Back to the example of analyzing a company’s finances. You might use a platform like Box to conduct research into all of the company’s unstructured data via agentic AI. But say you also want to turn that data into a digestible presentation for your team. In this paradigm, a presentation agent might call a secondary research agent, gather information, then use its own context engineering to provide the output you want, in the right format.
“This is common,” Kus says, “agents calling other agents. It’s the idea of an agentic ecosystem where all these different platforms provide their capabilities either as tools like an MCP, server context sense, or as agent-to-agent capabilities.”
In an agentic ecosystem, teams of agents cooperate like this in order to solve problems. Context engineering plays a pivotal role in crafting these dynamic ecosystems, ensuring that agents excel not just individually but in collaboration with each other.
The evolution of context engineering
“In the world we’re in now with this emergence of AI agents and models continuing to get better,” says Kus, “we’re seeing an evolution where many companies are making their agents more and more sophisticated using the practices of context engineering.”
This evolution isn’t just theoretical; it has dramatic potential to influence industries ranging from healthcare to manufacturing, education to finance. Enterprises that fail to embrace context engineering may find their AI struggling to deliver real value, while their competitors race ahead with smarter, more nuanced systems.
“Anyone who’s building AI agents needs to worry about context engineering,” Kus says.
Prompt engineering — still just as relevant
So does that mean prompt engineering isn’t important anymore? Not at all. The way you, as a human, communicate with an AI agent is still critical to the results you’ll achieve.
"You can actually give poor instructions to almost any agent,” Kus admits, “and it will not do what you want — because you didn’t communicate with it appropriately, even though it’s context-engineered really well."
But context engineering, which Ganesh quips is the “programmatic essence we put into AI agents,” is foundational to the success of agentic AI. If the agent is context engineered well, it will know how to ask for the information it needs in the first place.
Kus elaborates: “If you're going to get good outcomes out of your agents — especially the more complex they get — you’re going to need to focus on context engineering."
Catch the full episode
Context engineering has emerged as an essential approach for advancing AI capabilities. Moving beyond the era of simple prompt-based interactions, enterprises are adopting sophisticated frameworks to ensure AI systems perform complex, meaningful tasks.
For organizations, investing in the discipline of context engineering is non-negotiable. Whether you’re an engineer building internal AI tools or a business exploring AI solutions, the principle is clear: great outcomes rely on great context.
Ready to dive deeper into this discussion? Don’t miss this episode of the Box AI Explainer Series.Subscribe now and start listening today to learn practical, actionable strategies for integrating AI into your organization or industry.
