When someone joins a project, access to its files isn't the same as understanding how it works. The same is true for an agent. Code can show what the system does, while a design discussion explains why it works that way and a recent incident reveals where that understanding is incomplete.
Relevance matters more than volume
Providing everything the organization knows can make the important information harder to find. The context for a change should help answer the decisions that change raises: which behavior must remain intact, who depends on it, and what previous work has already established.
Relationships help. A work item linked to its implementation and review can explain more than either record alone. A dependency relationship can reveal whose work may be affected. Those connections reduce the amount a person or agent has to reconstruct, but they don't remove the need to judge whether the information is current and complete.
The absence of a result also needs interpretation. An inaccessible project, an outdated index, and a system with no dependency aren't the same situation. When that difference could change the decision, the next step may be gathering more information rather than proceeding with an apparently complete answer.
Context is maintained work
Instructions are most dependable when someone owns them and has a reason to revisit them. Build commands can stop working, an architectural decision can be superseded, and a workaround can become unnecessary. Repeated mistakes are often a reason to inspect the context before adding another instruction.
A project instruction file such as AGENTS.md can provide a short route into conventions and project knowledge. Its role is to help an agent find and use the right information, not to absorb every document the team has written. Keeping detailed guidance with the people who maintain it can make changes easier to follow, provided the agent can actually reach that guidance.
Context also includes what was learned during execution. A finding that changes the plan should be preserved with the work, while an enduring lesson may deserve a change to project guidance or tests. The team can then benefit from it after the original conversation ends.
Access doesn't make every statement an instruction
An issue, log entry, or external document may contain important facts alongside inaccurate claims or instructions written by someone who doesn't control the work. An agent needs enough access to investigate without treating everything it reads as authority to act.
Access controls and tool permissions provide enforcement that prose alone can't provide. The practical question is which information and actions the task requires, including where data will be sent when a model or external tool processes it. Removing unnecessary access reduces exposure, while excessively restrictive access can leave the agent making decisions with missing context. Both deserve attention.
Further Reading
- Making organizational context available to agents develops a way to identify missing information, connect sources, and maintain the context as work changes.
- AGENTS.md customization files explains where GitLab uses project instructions and where behavior differs between experiences.
- Security threats in agentic systems explains how untrusted content and tool access can interact.