What Is a Custom AI Agent Made Of?
See the parts every custom agent is assembled from, and how each one shapes what it does.
You’ve now met every ingredient. A custom agent, whether it’s a coding assistant, an expenses helper or a customer-support agent, is built from the same handful of parts. What makes it “yours” is how you choose and configure them.
- A model, run in a loop. Not always the biggest: a smaller, cheaper model is often enough for routine work (Track 5’s model choice).
- Instructions: its role, goals, rules and style, written as a system prompt.
- Tools: exactly the actions the job needs, and no more.
- Knowledge: documents it can search (Track 6), such as the expense policy.
- Memory: what it should remember between sessions.
- Guardrails: limits, approvals and checks that keep it safe.
Each part fixes a different failure. Without the policy, an expenses agent happily submits a claim that breaks the rules. Without a spending limit, it submits a €2,400 hotel bill without a manager’s say-so. Give it a tool it doesn’t need, such as deleting files, and one ambiguous request (“clean up my old stuff”) can turn into a disaster.
That last point is a good general rule: give an agent the smallest set of tools and permissions that lets it do its job. Every extra power is another way for things to go wrong.
Building it is the start, not the end. Like the prompts in Track 5, a custom agent is tested against realistic tasks, tricky ones included, and retested whenever its instructions, tools or model change.
Key takeaways
- A custom agent = a model in a loop + instructions + tools + knowledge + memory + guardrails.
- Each part prevents a different failure; missing ones show up as specific mistakes.
- Give the smallest set of tools and permissions the job needs.
- Test the agent on realistic and tricky tasks, and retest after every change.
Check your understanding
Go deeper (more technical)
Many platforms let people assemble agents without code: pick a model, write instructions, attach documents and connect tools, often via MCP. Developers use agent SDKs and frameworks that provide the loop, tracing and guardrails, and add their own tools.
A useful habit is to write the agent’s tests before building it: a list of tasks it must handle, tasks it must refuse, and tasks where it must ask a human. That list becomes its evaluation set (Track 5’s benchmarks lesson).
The simulation’s company and tests are made up.