Capstone Projects
Build four real agents end to end
Put it all together with four guided builds: a research assistant, a support agent, an autonomous coding pipeline, and a personal ops agent.
- Lessons
- 5
- Reading time
- 69 min
- Level
- Advanced
- Access
- Full course
Lessons in this module
- 11.117 min
Capstone 1: A Deep Research Assistant
Build a research agent that plans queries, reads sources, tracks citations, and writes a sourced report.
- 11.215 min
Capstone 2: A Customer Support Agent
Build a support agent with a knowledge base, account lookup tools, escalation rules, and an eval suite.
- 11.315 min
Capstone 3: An Autonomous Coding Pipeline
Wire up a pipeline where agents pick up issues, implement changes in isolated branches, run tests, and open pull requests for review.
- 11.414 min
Capstone 4: Your Personal Ops Agent
Build a scheduled agent that triages your inbox, summarizes your day, and drafts replies — with you in control.
- 11.58 min
Where to Go Next
How to keep your skills current as models and tools evolve, plus a roadmap of next projects to build.
About this module
The fastest way to consolidate what you know about agents is to build complete ones. A capstone project brings the separate pieces together, including a model loop, well-designed tools, context management, permissions, evals, and safety controls, all working on a realistic problem with real trade-offs.
Each project in this part is a guided build of a kind of agent people genuinely use, designed with the controls that make it trustworthy. The emphasis throughout is on keeping irreversible actions behind human approval or limits enforced in code, and on testing each agent with evals rather than spot checks.
The first capstone is a deep research assistant that plans queries, reads sources, records findings with exact quotes, and writes a Markdown report whose citations are checked by code. It runs on a mock web, so you need no extra accounts. The second is a customer support agent that answers from a knowledge base, looks up only the signed-in customer's account, issues refunds only within limits enforced in code, escalates to a human when the rules say so, and comes with an eval suite you run before every change.
The third is an autonomous coding pipeline that picks up issues, works in an isolated git worktree per issue under a strict permission policy, verifies its changes independently, and opens draft pull requests for human review. The fourth is a personal ops agent that triages your inbox each morning, drafts replies in your voice, and writes a daily briefing, while sending nothing without your approval. A final lesson covers how to keep your skills current and what to build next.
Questions about Capstone Projects
What are good AI agent projects to build?
Good projects solve a real, recurring problem and produce results you can check: a research assistant that cites its sources, a support agent backed by a knowledge base, a coding pipeline that turns issues into pull requests, or a personal assistant that triages email. Start with something you would use yourself.
How do you build a deep research agent?
Give the agent search and page-reading tools, have it plan queries, record findings with exact quotes and source links, and write the report from those notes. Checking citations in code against the saved quotes catches claims the sources don't actually support.
Can AI agents handle customer support?
They can handle many routine questions well when they answer from an up-to-date knowledge base and use tools scoped to the current customer. Policies such as refund limits should be enforced in code rather than in the prompt, and clear escalation rules should hand complex or sensitive cases to a person.
Can an AI agent turn GitHub issues into pull requests?
Yes, coding agents can implement well-scoped issues and open pull requests automatically. A safe pipeline isolates each task in its own branch or worktree, restricts what the agent can run, verifies changes with tests, and requires a human review before anything merges.
Should an AI email assistant send emails automatically?
For most people, no. Let the agent triage and draft, and keep sending as a separate step you approve, since a sent email can't be recalled. That captures most of the time savings while keeping the one irreversible action under your control.