Context engineering is the work of deciding what information an AI model or agent receives for a task: which facts, documents, history and rules, from which sources, inside whose permissions. For a company it can include governed context shared across supported tools, with appropriate access controls.
Prompt engineering and context engineering
This guide is for companies deciding how their people and AI agents get company context. If you are an engineer building context handling into your own agent, developer libraries, open-source projects and protocol servers for agent context are a more direct place to look.
Prompt engineering includes designing instructions and requests for a model.
Context engineering is about everything else the model sees when it answers: the instructions, the relevant facts, the history of the work, the tools it may use and the rules it must follow. Missing or stale context, poorly scoped permissions, unclear instructions and model errors can all cause problems.
What goes into context
For a task inside a company, the context an AI model needs usually has five parts.
- Instructions: what the task is and what a good result looks like.
- Company knowledge: the documents, decisions, meetings and records that bear on the task.
- History: what happened last time, and what changed since.
- Tools: what the AI may look up or do, such as search, a calendar or drafting an email.
- Rules: whose access the AI is working with, and what needs a person's approval.
Instructions and tools are usually set once. Company knowledge, history and rules are where the hard work is.
Why company context is hard
A model given the internet has too much context. A model given a company usually has too little, or the wrong version.
- It is scattered. The pricing decision is in a meeting, the exception is in a chat thread and the contract is in a drive.
- It changes. The terms that applied in March may not apply now, and a file rarely says so.
- Much of it was never written down. How the business actually runs often lives in a few people's heads.
- It is permissioned. A salary, a board discussion or a customer's contract should reach only some people, and so only some people's AI.
- It needs sources. An answer nobody can trace back to a record cannot be checked, so it cannot be trusted.
What governed context means
Governed context is company context that carries its rules with it. Four properties make it governed rather than merely connected.
- Current: dates and version history can help identify changes, but do not prove which fact is correct.
- Explained: decisions keep their reasoning, so AI can say why, not only what.
- Sourced: an answer names the record it came from, or says it has none.
- Permissioned: retrieval is scoped to the authorised user or service identity and the task.
These are the same properties described in the guide to AI legibility, seen from the engineering side.
The context layer
A company can configure context for individual tools, though separate integrations can create inconsistent copies and permissions.
One alternative is a governed context layer that serves selected information to compatible tools and agents. Open standards such as the Model Context Protocol, or MCP, make it easier for AI clients to read from one layer rather than from a dozen copies. The standard moves the context. It does not decide what the context should contain or who may see it, which is why the layer underneath still has to be governed.
Common context engineering mistakes
- Connecting everything at once. More context is not better context. Start with what one question needs.
- Treating every file as current. Old drafts and replaced policies give confident wrong answers.
- Leaving permissions for later. Access rules are far harder to add after AI has already read everything.
- Skipping the reasoning. Without the why, AI repeats past decisions without knowing when they stopped applying.
- Accepting answers without sources. If an answer cannot be traced, it cannot be checked.
About Hoppflow
Where Hoppflow fits
Hoppflow is a living company brain, the shared memory and governed context inside an AI operating layer for your company. It keeps company knowledge, meetings, documents, connected tools and operating records in one company memory, with recorded decisions and the reasoning behind them.
- Dates: recorded decisions and facts carry the dates when they applied and when Hoppflow learned them. Dates help review but do not prove correctness.
- Explained: recorded decisions keep their reasoning.
- Sourced: a company-brain answer names the record it came from, or says that it has no source.
- Permissioned: supported company-brain retrieval paths apply each person's access rules.
Connecting a supported tool does not make its contents searchable by default: indexing is a separate choice, except that a calendar shared with the workspace supplies meeting events. See the Company Brain page, the controls on Security and trust, or join the waitlist.
Questions people ask
What is context engineering in AI?
Context engineering is the practice of deciding what information an AI model or agent receives for a task: the instructions, the relevant facts and documents, the history of the work, the tools it may use and the rules it must follow, drawn from the right sources inside the right permissions.
How is context engineering different from prompt engineering?
Prompt engineering shapes instructions. Context engineering also considers the facts, history, tools and rules supplied to the model. Missing or stale context, unclear instructions and model errors can all contribute to wrong answers.
What is a context layer for AI?
A context layer can provide selected information and access controls to compatible AI tools. It is one possible architecture, not a requirement for every system.
What is governed context?
Governed context is company context that carries its rules with it. It can include dates, decision reasoning, source attribution and access controls for users or scoped service identities. These controls do not guarantee correctness.
Does MCP solve context engineering?
No. The Model Context Protocol is an open standard that lets AI clients connect to sources and tools. It moves context from one place to another. It does not decide what the context should contain, whether it is current, or who may see it. That still has to be governed.