Building AI Agents
Final Project Ideas For
Student-Built Agent Systems
Alex Avdiushenko
Fall 2026
Start with a real task
The project is the useful, testable system you build — not the number of agents.
01 · Goal
A real task
Who needs help, and what outcome should they get?
02 · Agency
Bounded decisions
What should the system decide, look up, or do with a tool?
03 · Evidence
Visible quality
How will you distingiush a successful run from a confident failure?
If one prompt reliably solves the task, keep it simple. Add an agent loop only when planning, tools, memory, or feedback improve the result.
Five directions to explore
Choose a narrow user and workflow; change the domain to make it yours.
Research copilot
Search a small document collection and return an answer with evidence and source links.
Workflow assistant
Turn a request into a plan across mock tasks, calendar events, or project tickets.
Language AI Tutor
Learn a new language. Use images, audio clips
and dialogs to help with a next level of new language.
Kaggle Prize Winner
Explore a provided dataset, metric and make AI win competition for you.
Review team
Use specialist roles to critique a draft, compare options, or check a report before it is delivered.
These are prompts, not prescribed topics. A focused version of your own idea is just as welcome.
Example: evidence-first research copilot
A useful answer should be traceable to the material it searched.
Workflow
Question → Retrieve sources → Draft with citations → Check support
Keep the scope small
Start with curated documents. Add retrieval, source snippets, and a clear “not enough evidence” response.
Make failures visible
Test answer correctness, citation support, retrieval misses, and questions the collection cannot answer.
A good demo includes one question the agent answers well and one where it should fail.
Example: workflow agent in a sandbox
Let the agent prepare actions; keep consequential changes under user control.
Possible task
Plan a busy week
Given mock calendar events and a to-do list, suggest a schedule and explain conflicts.
Tools
Read, draft, confirm
Read fixtures, create a proposed change, show a preview, then write only after explicit confirmation.
Evaluate
Task completion, valid tool arguments, conflict handling, and unnecessary actions.
Guard
Use fake data and a sandbox; never send messages, spend money, or change a real account in a demo.
Choose the smallest system that works
Architecture choices should follow the task — then be justified with evidence.
One agent first
Use a single loop with clear tools. Add specialist agents only if separation improves quality or makes the system easier to evaluate.
Memory with a purpose
Keep only state that helps across steps or sessions. Make its source and lifetime clear, and let users inspect or reset it.
Tools with boundaries
Give each tool a narrow job and observable inputs/outputs. Validate arguments and handle errors and timeouts.
Fallbacks are features
Ask a clarifying question, return partial results, or stop safely when evidence or permissions are missing.
Make quality measurable
Build a small, repeatable evaluation set before polishing the demo.
Task quality
Did it reach the intended outcome? Were answers correct and supported?
Agent behavior
Did it select the right tool, use valid arguments, and stop at the right time?
Engineering
What are latency, cost, and failure rates? Can you reproduce and inspect a run?
Include normal and ambiguous cases, missing data, tool errors, and attempts to exceed scope. Compare against a simple baseline and report wins as well as failures.
Your project pitch
A clear first version fits in six short answers.
The job and the agent
- User and job: Who is stuck, and what are they trying to accomplish?
- Agent loop: What decisions does the system make, and when does it stop?
- Tools and data: What may it read or change? What can you safely simulate?
The boundary and the proof
- Boundary: Which actions need confirmation, and how can the agent fail safely?
- Evaluation: What cases and metrics will show whether it works?
- Demo: What is the smallest end-to-end slice you can show?
Pick one workflow, make it reliable, and let the evaluation tell you what to build next.