After launch, the most important step is onboarding. An AI agent does not know you by default. To help it fit your tasks, get to know it as you would a new employee: ask it to pose clarifying questions with answer options. Your answers form a foundation for working together
Next, connect the tools it will use: email, calendar, a task manager and, when needed, MCP tools. The third important step is memory: decide what the agent should save, which rules it should follow and how often to review those records
Introduce the agent to your work
Explain what you do, which tasks you want to delegate, how you prefer to communicate and which topics matter. Ask the agent to offer answer options while leaving room for your own response. If the work involves clients, explain when it must stop and ask you before making a promise or sending a message
After the conversation, ask the agent to show you what it saved and check the wording. Correct any misunderstanding about your work or preferences before the agent starts relying on it
Connect the tools it needs
Start with the services required for the first tasks. A calendar can help find free time, email can help process messages, and a task manager can hold your to-do list. Connecting a service does not mean giving the agent unrestricted access: check what it can read or change and where it needs your approval
MCP is one way to connect AI applications to external tools and data. The MCP documentation explains how it works. Connect the tools your agent actually needs
Set up memory and review it
Decide which facts and instructions the agent should retain between conversations. I use rules that keep facts concise, link to more detail when needed and move files into an archive. You can review and rotate memory daily or weekly so the agent's working memory does not grow unchecked
Before moving records to an archive, check that important information will not be lost and the agent can still find the details. Memory depends on the system: a conversation alone does not guarantee that an agent will save new information or work better without review
Revisit the setup
As you work together, periodically update the agent and review its tools, memory and core files. For example, Codex uses AGENTS.md for instructions, while Claude Code has CLAUDE.md. Other systems may work differently - ask your agent which files and settings it uses
Onboarding, tools and memory are the foundation I suggest starting with. You can tailor them to your tasks when building an AI agent