What goes in
The first design decision is what data to include. Not everything should be in the environment. Include documents the agents need to reference: policies, templates, style guides, previous work, reference materials, and termbases. Exclude sensitive data that the agents should not see, outdated documents that would confuse the agents, and raw data that has not been cleaned or structured. The boundary between "in" and "out" is a design decision you make deliberately, not a default you accept.
How it is indexed
Your data can be indexed in two complementary ways. Semantic indexing converts documents into embeddings for fuzzy, meaning-based retrieval. Text indexing maintains documents as searchable plain text for precise, keyword-based retrieval. The choice depends on how the data will be used. Reference materials that agents search for general information benefit from semantic indexing. Termbases and structured data that agents search for exact terms benefit from text indexing. Most environments use both.
Agent rules
Each agent in your environment operates under rules you define: what it can access, how it should respond, what constraints it must follow, and what format it should produce. These rules are expressed as system prompts, but in an integrated environment they are more than one-off instructions. They are persistent configurations that shape every interaction. You might have one agent configured for drafting (creative, higher temperature, broad access to reference materials) and another for reviewing (critical, low temperature, access to quality criteria and style guides). The rules are the levers that let you tune the environment to your needs.
Access control
Not every agent should have access to everything. An agent that answers student questions should not have access to the answer key. An agent that drafts public communications should not have access to internal HR records. Access control in the environment works the same way as access control in any information system: you define who (which agent, which user) can see what (which documents, which indexes) and do what (read, generate, modify). A security measure and a quality measure simultaneously; it prevents data leaks and prevents the wrong context from contaminating the agent's output.
Examples
Separating draft and review agents
You create two agents for a writing workflow. The drafting agent has access to reference materials and is configured with a higher temperature for varied first drafts. The review agent has access to your quality criteria and style guide, is configured with low temperature for consistent evaluation, and has no access to the reference materials the drafter used. This separation prevents the reviewer from simply validating the drafter's approach; it forces an independent evaluation against your stated criteria.