Why not put every source into one prompt?

A model can work only with the information available in the current call. An organization may have correct knowledge, yet an agent can still make a poor decision if its input contains outdated documents, irrelevant records, or tool output without access context. Context engineering asks what this user and this task need now, and changes the supplied information as the task progresses.

How is context selected and updated?

A typical workflow identifies the user's goal, identity, and available resources, then selects material from conversation state, enterprise knowledge, business semantics, and tools. Retrieval can add evidence. The application can prioritize rules, compress long history, and revise the next input after a tool call. For example, a data agent first receives a metric definition and accessible data scope. After executing a query, it receives the result and execution status. Those steps need different context. How MIP helps AI understand your business shows how data, business meaning, and evidence can enter an analytics workflow.

Context engineering also manages conflicts and freshness. If two documents define a metric differently, the application should prefer a confirmed, currently valid definition and retain its source. Sensitive data must be protected when retrieving records and executing tools; a prompt telling the model not to disclose it is insufficient.

How does it differ from prompt engineering and RAG?

Prompt engineering primarily improves instructions and their wording. Context engineering also governs retrieved content, tool access, memory, task state, and update policies. RAG is one way to obtain external knowledge and can be part of context engineering, but it does not manage the entire agent workflow. A context window is the model's capacity for a single call; a larger window does not remove the need to select useful information.

FAQ

Is more context always better?

No. Irrelevant content consumes capacity and may conflict with valid rules. Prioritize relevance, provenance, and freshness.

Is context engineering only for agents?

No. Search assistants, summarizers, and coding assistants also need well selected inputs. Agents place greater demands on updates across decisions and tool calls.

How can it be evaluated?

Use real task examples to measure evidence retrieval, citation accuracy, task completion, permission compliance, latency, and cost. Inspect failures against the context actually supplied at each step.