What is context engineering?

Context engineering is the practice of deciding everything a language model sees at the moment it works: the instructions, the facts retrieved for the task, the results of earlier tool calls, what the agent remembers, and in what order and form. For an AI agent, the context is the whole of its view of the world, so building it well matters more than any single phrasing.

Why does context engineering matter once agents act?

An agent's decisions can only be as good as what is in front of it. If the refund policy was not retrieved, the agent will improvise one. If last week's case note sits next to today's ledger balance with no sign of which is newer, the agent may pick the wrong one. If an email from outside sits in the same block of text as the agent's instructions, the email can pass itself off as an instruction.

Each of those is a context problem, not a model problem. Swapping in a stronger model rarely fixes them.

How does context engineering work in practice?

It treats the context as an assembled product with parts that each have a job:

• Instructions: the role and rules, drawn from the agent's brief, kept separate from task data.
• Facts: only the records the task needs, each with its source, its provenance class and when it was read.
• Tool results: trimmed to what the next step needs, so a long response does not crowd out the policy.
• Memory: notes from earlier steps or tasks, labelled as the agent's own notes rather than as source facts.
• Untrusted content: inbound emails, attachments and web pages, fenced off and marked as external.

The size limit forces trade-offs. A context window holds only so much, so choosing what to leave out is half the work. For long-running tasks, deciding what to carry forward and what to fetch again is covered in how AI agents remember context between tasks.

What does that look like in an example?

A building society's mortgage servicing agent handles a request to overpay. A poorly engineered context dumps in the whole customer file, three years of statements and the entire product guide. A well-engineered one carries the product's overpayment terms, the current balance from the loan system with its timestamp, the overpayments already made this period, and the customer's message marked as declared. The agent answers faster, and a reviewer can see exactly which facts it was given.

How is context engineering different from prompt engineering?

Prompt engineering is about wording: how to phrase instructions so the model responds well. Context engineering is about assembly: which information goes in, from where, labelled how, and in what order, often across many steps and tool calls. Prompt engineering is one part of it. For agents that act over several steps, the assembly matters far more than the phrasing.

Where Pryme fits

The Pryme Context Protocol builds the facts part of an agent's context. Every fact arrives stamped with its source and provenance class, so the model is never asked to guess where something came from, and the platform can check the class again at the moment of action. See the governed agentic operating system, explained.