What Is Context Engineering? A Field Guide to the AI Context Window

Watch the full session · 23 min

When an AI agent gets a refund policy wrong, the model often never saw the policy. At COLLIDE 2026, Redis Senior Solution Architect Jessica Emmons explained context engineering, the four ways context fails, why basic RAG isn't enough, and how context engines, agent memory and semantic caching fill the gap.

Imagine a returns agent telling a customer they qualify for a refund within 60 days when the real policy is 30. It is easy to blame the model, Jessica Emmons, Senior Solution Architect at Redis, told attendees at COLLIDE Data + AI Conference 2026 in Atlanta. But "the model never saw your refund policy," she said. The failure happened upstream, in what was loaded into the context window.

In her session, "What's in the Context Window: A Field Guide to Context Engineering," Emmons explained what the practice is, why retrieval-augmented generation (RAG) alone falls short, and which architecture patterns she sees working.

What is context engineering?

For the last few years, the industry focused on model intelligence, Emmons said. Once companies started trusting agents to take actions, the question shifted from why the model got it wrong to what the team gave the model.

She cited Andrej Karpathy's definition: "the delicate art and science of filling the context window with just the right information for the next step." Redis defines it as deciding what goes into the context window at each step of an agent's run, including system instructions, tool definitions, tool call results, retrieved documents and working state.

The guiding principle, she said, is to "find the minimum number of high-value tokens in order to maximize the probability of a successful result."

Why isn't prompt engineering enough for AI agents?

A well-written prompt works for a single-turn chatbot, Emmons said. Agent workflows run through many steps and call many tools, and tool outputs pile up in the context. Over several steps that context can outgrow the window, which raises latency and cost and degrades reasoning.

Without a strategy for what to keep, retrieve, compress and discard, she said, teams risk agent drift, forgotten information and hallucinations.

What are the most common context failures?

Emmons grouped them into four categories:

• Fragmented data: context spread across many systems with nothing linking them, so the agent never sees the refund policy.
• Data opacity: the agent can reach the data but doesn't understand relationships, such as different refund policies for different item types.
• Limited accumulation: new information from users doesn't persist, so the agent doesn't improve.
• Speed degradation: slow steps add up. She said one customer was considering scrapping an AI project because each conversation turn took up to six seconds.

A Redis survey of about 1,000 enterprise AI leaders in August found that 97% agree context is very important to success, but only 4% have built for it, Emmons said.

Why does RAG fall short for enterprise agents?

Basic RAG is linear: a question triggers a vector search, and the results go to the model. That leaves no room for the model's reasoning abilities, such as rewriting a prompt or retrieving again, Emmons said. It also relies on stale documents rather than live operational data.

Ask a RAG-based returns agent why a refund hasn't arrived, and it lists generic reasons and a support email. "This is not an inaccurate answer, it's just not a helpful answer to the user," she said. A bigger vector database doesn't fix that, according to Emmons.

Connecting the agent directly to every back-end system isn't the answer either. "Congratulations, you basically made the fragmentation problem in the agent itself," she said, and a model that sometimes does unexpected things is now wired to systems of record.

What is a context engine?

Emmons described a context engine as a single semantic and access layer between agents and back-end systems. It holds structured data definitions, unstructured data for RAG, and short- and long-term memory, kept in sync with the source systems. Schema-based, governed MCP tools let agents navigate it.

That design lets an agent answer about a specific order at a specific time, she said. Short-term memory covers session state and chat history. Long-term memory carries preferences and past orders across sessions, so "your 10th time using an agent is going to be better than the first."

For speed and cost, she recommended two patterns. Semantic caching answers semantically similar queries from a cache instead of calling the LLM, which she said can cut token costs by up to 70% at enterprise scale. Semantic routing adds deterministic logic that blocks irrelevant queries, sends simple ones to cheaper models and escalates when needed.

The storage layer still matters, Emmons said, because one agent may need key-value lookups, semantic search, exact matches and hybrid search within a single latency budget. "If any of this is slow, the whole thing's slow."

She said Redis has packaged this architecture as Redis Iris, its real-time context engine, while acknowledging teams can also build their own. The goal, in Redis's maturity model, is "compounding" context, where each new agent in production makes the shared context better instead of being built in a vacuum.

Context engineering at COLLIDE 2027

COLLIDE, the largest AI conference in the Southern U.S., returns Wednesday, October 13, 2027, to the Sandy Springs Performing Arts Center in Atlanta, with more than 1,000 enterprise data and AI leaders expected. Learn more about COLLIDE, or see hands-on sessions in Data Science Connect's free 2027 webinars, including AI Demo Days.

Watch: What's in the Context Window: A Field Guide to Context Engineering on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights