In this article, we'll cover the current state of context engineering, how it powers modern adaptive experiences and how we’re applying these concepts at Contentstack to streamline our own operations.
We all know that the quality of an LLM’s output is entirely dependent on the quality of its inputs. Give an AI tool a lazy prompt, and the result is usually slop.
But if you really want an AI agent or custom LLM to do incredible things, you need to provide more than a detailed, carefully worded set of instructions. You need to arm that AI tool with context.
Context engineering is a critical element of AI development focused on giving LLMs everything they need to solve complex problems and provide true value to business teams and consumers.
In this article, we’ll cover the current state of context engineering, how context engineering powers modern adaptive experiences and how we’re applying these concepts at Contentstack to streamline our own operations.
What is context engineering and why does it matter?
As Google DeepMind’s Phil Schmid defines it, context engineering is “the discipline of designing and building dynamic systems that provide the right information and tools, in the right format, at the right time, to give a LLM everything it needs to accomplish a task.”
Context engineering is necessary because LLMs often require more than a simple text prompt to generate a useful response. (To use a construction metaphor: You need more than a blueprint to build a house. You also need a variety of tools, relationships with electricians and plumbers, and a good source of lumber.)
Context for an LLM or AI agent could mean:
- Access to relevant data, knowledge bases and retrieved information
- A historical record of the instructions you’ve previously given it
- Explanation of all the tools the agent can use to perform tasks
- Structured output, or your specific preferred format for the LLM’s response
- Instructions that define the preferred behavior of the LLM while it’s being used
- A user prompt that sets the LLM off to accomplish its task(s)
In other words, context is all the information that an agent needs to know before it can intelligently respond to a prompt or request. Context engineering is the process of providing that information.
“The main thing that determines whether an Agent succeeds or fails is the quality of the context you give it. Most agent failures are not model failures anymore, they are context failures.” — Phil Schmid, Staff Engineer, Google DeepMind
Examples of context engineering in the real world
To dig into this topic further, I spoke to Aniketh Shenoy Kota, a computer science engineer specializing in AI at Contentstack who has been exploring context engineering for the better part of a decade.
As Aniketh explains, a simple example of context engineering can be seen in AI writing assistants that know just enough about a situation to be useful.
“Imagine you ask an AI to ‘help me reply to this customer about their delayed shipment,’” Aniketh says. “Prompt engineering would focus on the literal wording of that sentence. Context engineering is everything around it: the system also has the original email thread, knows who the customer is, can look up their order status, sees that this is their third delay this year and understands your brand’s tone-of-voice guidelines.
“When that context is wired in properly, you don’t need a perfect prompt. You can say something natural, and the AI still produces a response that’s accurate, empathetic, on-brand and operationally correct; for example, offering the right make-good. That’s context engineering: giving the model the right information and tools, in the right format and at the right time, so the task is actually solvable — not just eloquently worded.”
A more complex example of context engineering could be an LLM-powered “customer navigator” that helps a SaaS company understand what’s really happening across an account. On the surface, you might ask the model, “Explain why Company X’s renewal risk has gone up and what we should do next.”
Under the hood, context engineering is doing the heavy lifting. The system has to:
- Unify structured data from CRM, product analytics, billing and support tickets
- Ingest unstructured signals from call transcripts, emails, Slack or Teams threads and survey comments
- Normalize everything into a shared semantic model (accounts, stakeholders, incidents, milestones, outcomes)
- Retrieve the right subset of that universe into a compact, well-structured context that an LLM can reason over
“Once you’ve engineered that context,” Aniketh says, “the model can do something genuinely advanced: generate a defensible churn hypothesis, highlight the specific interactions that moved the risk up or down and recommend next-best actions tied to proven playbooks. None of that comes from a clever one-line prompt. It comes from designing the whole environment in which the LLM operates.”
How important is context engineering to adaptive experiences?
The ability for brands to gather context from their website visitors and adapt to their needs in real-time is what will separate the winners from the losers in the years to come. (“Context is queen,” as we like to say around here.)
For adaptive digital experiences, context engineering isn’t a nice-to-have, it’s the essential core of the architecture.
Traditional rules-based personalization takes a handful of attributes (“if user is in segment X and on page Y, show banner Z”) and calls it a day. That might work for simple targeting, but it’s brittle and quickly explodes in complexity.
Reasoning-based personalization, by contrast, asks: “Given everything we know about this visitor — their history, interests, current behavior, environment and our brand rules — what is the best thing to do right now?” And you can only answer that if you’ve engineered rich, real-time context around every interaction.
That’s exactly the shift our Agent OS release was designed around. Instead of treating content, data and “the AI” as separate islands, Agent OS turns them into a context fabric: agents can see customer behavior, brand rules, content variants and business goals in one place, and then reason their way to the next best experience.
This enables brands to evolve from content management to context management — from publishing static experiences to orchestrating live, one-to-one conversations with your audience powered by first-party data.
Remember, agents are only as smart as the context we give them: what data they can see, how that data is semantically modeled, which tools they can call and how we constrain outputs to stay on-brand and compliant. That’s what allows a reasoning-based system to go beyond segments and rules and actually adapt in the moment.
How Contentstack is using context engineering to improve internal operations
So how can context engineering make a global enterprise like Contentstack more effective? I asked Aniketh to give me a sneak preview of some of the internal efforts he’s working on:
“At Contentstack I’ve been focused on building a context-aware internal agent that’s a good microcosm of what Agent OS can do for customers.
The agent’s job is to understand customer sentiment and churn risk across a very fragmented landscape: Salesforce, Gong, Jira, Confluence, Slack, Google Drive and other business systems. Each of those sources speaks a different ‘language’ — structured objects, tickets, transcripts, documents, chat messages – so the first step is context engineering: We build a semantic layer that normalizes everything into a shared ontology of accounts, contacts, interactions, risks and outcomes.
“On top of that, we’ve designed a parallel sub-agent architecture. Specialized sub-agents focus on different modalities and sources (for example, one that reads Gong calls and scores sentiment and themes, another that scans Jira and support data for incident patterns, another that looks at product-usage anomalies). They each retrieve the most relevant snippets from their domain and annotate them with structured signals.
“A coordinating agent then pulls those pieces into a single, tightly scoped context window for each account and runs higher-level reasoning: ‘Given all of this, what’s happening with this customer, how risky are they and why?’
“The result isn’t just a score; it’s a narrative with citations back to the underlying evidence — specific calls, tickets, comments or product metrics that justify the assessment. We’re planning to feed that directly into our Customer Health Index so CSMs and leadership get a living, AI-curated picture of risk and opportunity instead of static dashboards.
“If you zoom out, the same pattern is what Agent OS can offer our customers. Instead of forcing teams to stitch together dozens of tools and rules manually, you get:
- a shared semantic layer across your content and data,
- agents that specialize by channel or task (web, mobile, email, commerce, service),
- and a reasoning layer that can adapt experiences in real time based on that unified context.
“That could mean an experience engine that notices frustration in support interactions and automatically dials down promotional messaging, or a storefront that dynamically rewrites content and recommendations based on a customer’s current mission rather than their static segment.
“All of that is powered less by ‘magic prompts’ and more by careful context engineering — which is why I’m so excited about where this work is going.”
