Where to Go Next
How to keep your skills current as models and tools evolve, plus a roadmap of next projects to build.
You've done something most people who talk about AI agents haven't: you've built four of them end to end, with evals, guardrails, and humans in the right places. You know how the loop works, where agents fail, and how to tell the difference between a demo and something you'd trust with real work.
The tools will keep changing, probably faster than any course can track. This last lesson is about what to do with that. You'll revisit the goals you set in the first lesson, set up a light routine that keeps your skills current without chasing every announcement, and pick your next projects from a roadmap built on what you've already made.
Revisit the note you wrote on day one
In the very first lesson, you wrote down three things: a task you do every week that you'd love to hand to an agent, something you'd like to build, and your biggest doubt about agents. Find that note now. It's worth reading before anything else in this lesson.
Your weekly task
Ask yourself honestly: has it been delegated? If not, you now have everything you need, and the right starting point depends on the task's shape:
| If your weekly task is... | Start from |
|---|---|
| Gathering information and writing it up | Capstone 1, or a chat agent with the research playbook |
| Answering the same questions repeatedly | Capstone 2's knowledge base and escalation design |
| Small code changes, fixes, or upkeep | A coding agent, then Capstone 3 once you trust it |
| Inbox, scheduling, or recurring reports | Capstone 4 and Scheduled and Background Agents |
| Cleaning or analyzing spreadsheets | Data Analysis with Agents |
If the task still feels too fuzzy to delegate, that's usually a task-definition problem, not an agent problem. Write it up using the goal, context, constraints, and definition of done from Writing Tasks an Agent Can Actually Finish, and it often becomes delegable on the spot.
The thing you wanted to build
Compare it with the four capstones. Most agent projects are a remix of their parts: a workflow with an agentic middle (Capstone 1), tools scoped to a user with policy in code (Capstone 2), an agent harness with a permission policy and independent verification (Capstone 3), and a read-draft-approve split with a human holding the irreversible action (Capstone 4). Name which parts your project needs, and you have most of an architecture.
Your biggest doubt
Here's where the course tried to answer the common ones. If yours is still open, the linked lessons are the place to go back to:
- "Is it reliable?" Reliability isn't a property of a model; it's something you measure for your task. That's what Why Evals Are Your Superpower and the rest of that module are for.
- "Is it affordable?" Cost is controllable with model choice, caching, batching, and budgets. See Controlling Cost and Latency.
- "Is it safe?" Not by default, but it can be made safe enough for many jobs through design: least privilege, sandboxing, and human approval for irreversible actions. The Safety, Security & Guardrails module covers the playbook.
- "Is this hype?" You've now seen the real limits up close: agents that wander, invent, or follow injected instructions. You've also seen them do hours of useful work when the task, tools, and checks are right. You don't have to take anyone's word on this anymore, including this course's.