Roadmap
AI agents
Tool calling, agent loops, MCP and safety: agents that do real work without going rogue.
Who it's for
Developers who have built with LLMs and want them to take actions, not just answer.
Where it leads
- AI engineer
- Agent developer
- Automation engineer
By the end you can
- Know when an agent is the right tool and when it isn't
- Build an agent that uses tools reliably
- Connect tools through MCP
- Design approvals and limits so an agent can't do damage
- 1
What an agent is, and when not to build one
Most problems want a simple workflow. Learn the patterns before reaching for autonomy.
- 2
Learn by building
A structured, free course from a single tool call to a working agent.
Do this: Build an agent with three tools that books, reschedules or cancels something in a fake system.
- 3
Connect tools with MCP
One standard way to give any agent access to tools and data.
Do this: Write a small MCP server for a tool you use and connect it to an MCP client.
- 4
Frameworks for real agents
State, memory, branching and multiple agents. Pick one and go deep.
- 5
Keep agents safe
Agents read untrusted content and take actions. Prompt injection is the threat to design around.
Do this: Add an approval step before every action that spends money or sends a message, and try to trick your own agent past it.
- 6
Evaluate the whole trajectory
Judge the steps an agent took, not just where it ended up.
Learning this with others is easier.
Ask questions when you're stuck, find people on the same path, and track your progress on DevLearn.