Georgia State University — J. Mack Robinson College of Business PATH — Pathways for AI Training & Hiring CIS 4394 Agentic AI  ·  Fall 2026  ·  Dr. Xinyu Fu
CIS 4394 · Week 2 ⌂ Course Week 2 Summary Answer key GA1 groups Tutorials Quiz 1
After class · Week 2

Week 2, wrapped — and what’s next.

Four things on this page: the full answer key for the repo scavenger hunt, how to join a group for Group Assignment 1, clear step-by-step tutorials for each build path, and a heads-up on Quiz 1 with sample questions.

Scavenger hunt · full answer key

The nine layers, mapped to real Codex files

Finding evidence in a 111-module repo was genuinely hard — that was part of the point. Here is the full map: one concrete file per layer, with what you’ll find inside. Remember: the repo has no folders named “Goal” or “Loop” — the nine layers are our architectural interpretation, not OpenAI’s official taxonomy. (Paths verified Sept 2026; this repo moves fast.)

LayerFile (click to open)What you find there
01 Goalmcp-server/src/codex_tool_config.rsThe goal is surprisingly thin — just “the initial user prompt to start the Codex conversation.” No GoalManager class. Persistence comes from keeping that objective in context across many actions. Prompt = one message; goal = that objective across a sequence of actions.
02 Model / Reasoningcore/src/config/edit.rsSetModel { model, effort } — the model and its reasoning effort are configuration. The model decides what to do next; the runtime carries it out. Reasoning is a component, not the whole system.
03 Instructionsconfig/src/config_toml.rsInstruction layers beyond your prompt: “Developer instructions inserted as a ‘developer’ role message.” Goal = what outcome; instructions = how to behave while pursuing it.
04 Toolscore/src/tools/handlers/plan_spec.rsname: "update_plan" — a tool defined in code, with a schema (steps, statuses: pending / in_progress / completed) the model can see. Tools turn “input → text” into “reason → choose tool → act → observe.”
05 Environmentcore/tests/suite/prompt_caching.rsCodex injects environment context — the tests check for <cwd>, <shell>, <current_date>, <timezone> tags. The environment is the world the agent can observe and act on.
06 State / Memorycore/config.schema.jsonHistory persists to ~/.codex/history.jsonl, and a separate memories subsystem exists. Three kinds: turn state → conversation state → long-lived memory. You don’t need all three to be an agent.
07 Agent Loop ⭐core/src/session/turn.rsRead this file first. Its comments state the whole algorithm: if the model requests a function call, execute it and send the output back; if it sends only a message, the turn is complete. That’s the agent loop — and its stop condition — in two sentences.
08 Guardrails / HITLcore/src/exec_policy.rsDecision::Prompt → NeedsApproval — policy code sits between reasoning and action: allow, forbid, or ask the human. Model intention ≠ execution permission. This is runtime enforcement, not a polite prompt line.
09 Evaluationprotocol/…/base_instructions/default.mdThe base instructions tell the agent to verify its work with the project’s own tests and builds. Evaluation doesn’t need another LLM — pytest passing is evaluation. And a failing test feeds the loop: fail → observe → reason → retry.

The whole architecture, in one loop

while not done: response = model(goal, instructions, environment, history, tools) if response.requests_tool: if policy(response.tool_call) == NEEDS_APPROVAL: # layer 8 wait_for_human() result = execute(response.tool_call) # layer 4/5 history.append(result) # layer 6 else: return response # stop condition

The real Rust adds streaming, cancellation, compaction, errors, parallel tools — but architecturally, that is Codex. The single takeaway of Week 2: the LLM is only one component of the agent.

Want to read the repo yourself? Go in this order: core/src/session/turn.rs (the loop) → core/src/tools/ → protocol/…/base_instructions/default.md → core/src/session/turn_context.rs → core/src/exec_policy.rs → config / history / memories. Loop first — or you’ll drown in infrastructure.
Action required

Join a group for Group Assignment 1

GA1 (the Job Search Agent) is submitted as a group, and you must self-enroll in a group on iCollege — in your own section.

How to enroll (2 minutes)

  1. Open your section’s course on iCollege ↗.
  2. Go to Assessments → Groups.
  3. Click View Available Groups.
  4. Find a group with space and click Join Group.
  5. Confirm your name appears in the member list.

The rules

  • 1–6 members. Working solo is allowed — you still must join a group (of one).
  • Join the group you mean to join. Coordinate with your teammates first — don’t drop into a group of strangers, and don’t “save seats” in multiple groups.
  • Your own section only. Groups don’t transfer between sections.
  • Deadline: GA1 is due Tue Sept 15 (Wednesday section) / Wed Sept 16 (Thursday section), 11:59 pm EST — enroll well before then.
Build paths

Path tutorials — clearer this time

In-class was a quick look; here is the sit-down version. Each path: numbered steps, an official illustrated guide, and a video. Full lab requirements stay on the Build page.

🅐 Google Opal — the visual path

  1. Use a personal Gmail — your GSU account is blocked from Labs. You must be 18+.
  2. On a computer, open gemini.google.com ↗ → left sidebar → Gems.
  3. Under My Gems from Labs, click New Gem — this opens the Opal editor (boxes connected by arrows; each box is a step).
  4. Paste the full starter prompt from prompts/opal_starter_prompt.md (starter kit).
  5. Add three inputs by pasting the files as text: resume.md (facts), preferences.md (constraints), data/jobs.json (jobs).
  6. Build the flow: an evaluate step → routing (recommend / investigate / down-rank / ask) → a user-input step for Approve / Edit / Reject → a draft step that only runs after approval.
  7. Click Preview to run; test with J001 first, then all four cases. Screenshot every run — those are your evidence.

Illustrated official guide: developers.google.com/opal ↗ · account requirements: Google Help ↗

Video walkthrough (12 min)Building an agent workflow in Opal via the current Gemini → Gems route (AsapGuide, May 2026).

🅑 LangSmith Fleet — the builder path

  1. Sign in at smith.langchain.com ↗ (free account).
  2. In the left navigation, click Switch to Fleet.
  3. Click + New Agent → choose Create manually (skip templates for this lab).
  4. Paste prompts/fleet_starter_prompt.md into the agent instructions.
  5. Paste resume.md, preferences.md, and data/jobs.json into the agent’s context.
  6. In settings, require approval before final external-facing drafts. Do not connect email or any submission tools.
  7. Run the four cases from the chat panel. After each run, open the trace — that trace view is exactly the evidence deliverable.
  8. Budget: free tier ≈ 5 Fleet LCUs/month and runs pause when it’s gone — plan your runs, don’t spam retries.

Illustrated official guide: Fleet docs ↗ · free hands-on course: LangChain Academy ↗

Official: Introducing Fleet (7 min)LangChain, Mar 2026 — what Fleet is and the core flow.
Official: hands-on getting started (7 min)LangChain — recorded under the old name “Agent Builder”; same product, some labels have changed.

🅒 Code path — recommended: Google ADK

For those coding GA1, we recommend Google’s Agent Development Kit (ADK) — an official, open-source Python framework that runs on the same free Gemini API key you created in Week 1, and ships a built-in dev UI with a trace view (instant evidence for deliverables #8–9).

pip install google-adk adk create job_agent # scaffold # put your GOOGLE_API_KEY in job_agent/.env adk web # dev UI + traces at localhost:8000
  1. Follow the ADK quickstart ↗ to a running hello-agent.
  2. Give your agent tools to read resume.md, preferences.md, jobs.json from the starter kit.
  3. Implement the decision policy (recommend / investigate / down-rank / ask) and an Approve / Edit / Reject pause before drafting.
  4. Run the four cases; export traces from adk web.

Also fine: the plain-Python starter-kit baseline + any coding agent (Codex, Copilot, Claude Code). We’ll meet LangGraph — the course’s main framework — in Week 4; ADK is the fastest on-ramp for GA1.

Official: Introducing ADK (9 min)Google for Developers.
Official: getting started (19 min)Google for Developers — from install to a working agent.
Next week

Quiz 1 — foundations, Weeks 1–3

Five multiple-choice questions, closed-book, at the end of next week’s session. Everything comes from the assigned readings below. Read them; the quiz will feel easy.

W1 · Anthropic — Building Effective AI Agents
Workflows vs agents; the five workflow patterns; “simplest thing that works.” W1 · Bornet — Agentic Artificial Intelligence (LinkedIn Learning)
Free with your GSU login — but sign in through the GSU portal first (next link), or you’ll hit a paywall. ↳ GSU LinkedIn Learning portal
Sign in here with your CampusID once, then both LinkedIn Learning readings open free. W2 · Codex’s system prompt (gpt_5_2_prompt.md)
AGENTS.md rules, planning, task execution, validation. W2 · OpenAI Cookbook — Planning for Long-Horizon Agents
How Codex plans multi-step work. W2 · Anthropic — Claude Code auto mode
Autonomy and safeguards in a competing coding agent. W3 · Sheer — Introduction to Prompt Engineering (LinkedIn Learning)
Via the GSU portal, same as above. W3 · Hugging Face Agents Course — Unit 1
Free; what an agent is, in the HF framing.

Five sample questions

Same style and difficulty as the real thing. Try before revealing.

← BackWeek 2 home