Context engineering

Context engineering is the work of making sure an AI model has the right information, instructions and tools in front of it for each step of a task. Where prompt engineering is about how you word a request, context engineering is about everything else in the context window: the relevant documents, the conversation so far, the results of earlier tool calls, the company’s conventions and the tools available.

Why it matters

Most bad AI output comes from missing or wrong context, not a weak model or a badly phrased prompt. A model asked to draft a renewal email does far better when it already has the account history, the pricing rules and the tone the team uses.

Agents make this the main job. An agent working through a long task builds up context with every step, and its context window fills with tool output, old messages and dead ends. Deciding what to keep, what to summarize, what to fetch fresh and what to drop is what separates an agent that finishes the job from one that loses the thread halfway.

What it involves

  • Retrieval: pulling in the documents and records the task needs, and leaving out the rest.
  • Standing instructions: system prompts and files like AGENTS.md that tell an agent how this team or codebase works.
  • Memory and compaction: carrying forward what matters from earlier work without dragging the whole history along.
  • Tools: giving the model the right tools, with clear descriptions, and not dozens it doesn’t need.

More context is not better context. Overloaded windows cost more and perform worse, a problem known as context rot.

Context across a team

Context engineering is usually discussed for one agent. On a team, the harder problem is that each person’s agent has its own context, and what one learns never reaches the others. Shared context, in the form of threads, files and decisions the agents can all read, keeps them working from the same picture.

How elvex relates

elvex Spaces hold a team’s threads, files and to-dos, and each person’s agent reads them with that person’s access, so agents from different people start from the same context.

Bring your agent. Bring your team.