How to configure the agent
Configuration lets you give the agent context, pin down instructions, choose tools and restrict access. Do not put the same instruction into every mechanism: pick the one whose scope matches your task.
| Task | What to use |
|---|---|
| General project context and conventions | AGENTS.md |
| An instruction for a project or for files matching a glob pattern | Rules |
| A repeatable specialized task | Skills |
| Knowledge shared across chats | Memory Bank |
| The agent's role and the tools available to it | Agent modes |
| A model per type of subtask | Model profiles |
| External services and tools | MCP |
| Limiting the scope of the current edits | Edit Scope |
| Forbidding the agent to read or write files | .readignore and .writeignore |
| Tool access and confirmation prompts | Tools menu |
| Data transfer, local data and legal documents | Data, privacy and security |
| A setting does not load or conflicts with another | Configuration troubleshooting |
Global Rules are available in all of your projects, local ones only in the current project. A glob pattern additionally restricts the files a rule applies to.
Workflows are no longer part of the recommended path. Create Skills for new repeatable scenarios. If you still have old Workflows, follow the migration guide.
Where to start
- Add only the context that every task in the project needs to
AGENTS.md. - Create a Rule when an instruction has to fire for specific files.
- Capture a repeatable task as a Skill and run it manually with
/, or let the agent pick it automatically. - Before making changes, limit the scope with Edit Scope or
.writeignoreif the agent must not touch part of the project. - Enable Memory Bank, model profiles and MCP only when your process needs them.
How the mechanisms combine
Text instructions (AGENTS.md, Rules, Skills and memory) guide the agent but do not grant it any extra technical rights. Whether it can read, edit, run the debugger or call an external service depends on the mode, the available tools, Edit Scope, the ignore files and operation confirmations.
The current documentation does not define a single conflict resolution order across all kinds of instructions. Do not create contradictory rules. If the behavior is unexpected, start a new chat and temporarily keep the questionable instruction in one place only.
Exact details about scope, activation, known priorities, defaults, limitations and diagnostics are collected in the configuration reference.