CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
Four things on this page: a recap of what you should now be able to do, how to self-enroll in a group for Group Assignment 2 (new groups — and I want to see new names together), the build paths you can choose from, and the proposal that is due next week.
The model proposes a tool call; your harness disposes. Everything this week followed from taking that one sentence literally — because if your code is the only thing that executes, your code is the only place controls need to live.
| Page | The one thing to remember |
|---|---|
| 01 · Function calling | A “tool call” is two objects in order: the schema you write first (name, typed parameters, description), and the call the model emits second — a small JSON block naming a function and its arguments. The model never executes. Your harness parses, validates, runs, and appends the return value as the next observation. |
| 02 · Schemas, errors & gates | A rule in the prompt is a request. A rule in the schema is a wall. Three rails: least privilege (expose the narrowest tool that does the job), constrain the inputs (enum, min/max, required), separate duties (never one tool that both decides and pays). Then sort every tool by blast radius: auto-run · constrain · human-gate. |
| 03 · Build it locally | You ran a real agent loop on your own laptop, free, with no API key. A 0.5B model can call a tool correctly and still fail the task — and the fix was never “a bigger model.” It was a coarser tool, a tighter docstring, and a negative rule. Tool design is the lever. |
| 04 · Practice & case | Gate the effect, not the input. Reading hostile data is survivable; acting on it is not. And gate selectively — a gate nobody reads is not a control, it is a rubber stamp. |
GA2 is released on iCollege — build a tool-using agent: a loop, at least two tools with real schemas, a max-iteration guard, one clean execution trace, and a one-page design note. It is submitted as a group, and you must self-enroll in a group yourself.
Your GA1 group does not carry over. The GA2 groups on iCollege are empty and separate, so everyone enrolls again from scratch. You may re-form your GA1 team if you want to — but I would rather you didn’t.
Work with at least one person you have not worked with yet. Two selfish reasons, not just a nice sentiment: (1) GA2 is the assignment where skill mix decides the grade — the design note and the failure-mode analysis are worth as much as the code, and a team of all-coders or all-non-coders usually loses points on the half they are weaker at; (2) this is the last group assignment before you commit to capstone teams, so it is a low-cost way to find out who you actually want to build with for the rest of the semester.
GA2 is graded on agent design, not on which framework you typed it into. Both official paths are held to the same standard: no-code is not graded easier, and it does not excuse you from the design thinking. Pick one path per team and name it in the first line of your design note.
Use whatever you are fastest in:
Deliver: the code, a README with exact run steps, sample inputs. No API key in the repo.
If your team would rather not write code, build the same agent in a visual builder — LangSmith Fleet ↗ (free tier) is the supported default, and you met it in Week 2.
The requirements are identical: two tools, a loop with a stop condition, one trace, the design note. You will not be graded easier for going no-code — and you cannot use it to skip the tool/gate/failure-mode thinking.
Deliver: the exported flow or a shareable link, plus screenshots of the tool config and the loop/guard settings.
Google ADK, OpenAI Agents SDK, CrewAI, n8n, Opal, Copilot Studio, a local model through some other runner — all fine. If you already have a tool you enjoy, use it.
Two conditions, and they are the whole catch: it must be free (if it asks for a card, you are on the wrong plan), and it must let you show the five required things in the box below. Some pretty builders hide the loop and will not give you a real trace — check that before you build, not the night before it is due.
Not sure whether your tool qualifies? Ask me — one message, before you invest a weekend in it.
Full instructions and the point-by-point rubric are on iCollege — if anything here differs, iCollege wins.
The proposal is due next week, submitted on iCollege in your own section. The two sections’ deadlines are a day apart, so check iCollege for your section’s due date. The official prompt and rubric live on iCollege — this is just a last pass before you submit.
The proposal is individual to your capstone, not to GA2. They are different deliverables on different clocks — do not let the proposal deadline hide GA2, and do not let GA2 eat the proposal. Both are live right now.
Narrow beats ambitious. Every semester the strongest capstones are the ones that scoped down early. A refund helper that really works beats a “universal business assistant” that demos once and falls over.
A fuller checklist is on the Week 6 practice page →
This week you built a tool boundary by hand: schemas you wrote, validation you wrote, gates you placed. Next week that hand-built boundary meets the protocols designed to standardise it — so any tool can plug into any agent, and agents can talk to each other.