Autonomous Agents · Agentic AI for working professionals

Give AI a goal.
Build the Agent that follows through.

Before week 1 — what you need

  • R1Basic programming
  • R2Bring a laptop
  • R3Active ChatGPT Plus subscription + working Codex access

Build an Agent that uses tools, checks results, and knows when to stop. Turn your work experience into the judgment that guides its next move.

Fall 2027 · Academic year 2027–28, semester 1 16 weeks × 3 hours · Weekly time: to be announced

Fig. 01 · The Agent loop — step-through walkthrough

STEP 00 / 12

Idle — press the button to step through three full cycles.

    Sec. 01 · Why this course For working professionals

    Your expertise.
    An Agent that acts on it.

    A language model answers a question. An Agent pursues a goal — calling tools, checking its own work, deciding what to try next, and knowing when to stop or hand off to a person.

    Businesses are beginning to build their own AI tools around the work they know best. Your experience helps you choose a useful task, define a good result, and decide when a person should take over.

    AGENT
    A program where the model decides actions step by step, instead of answering once.
    TOOL CALL
    Letting the model run a function you wrote — and read the result.
    AGENT LOOP
    Decide → act → observe → repeat, with a step limit and a stopping rule.
    EVALUATION
    Scoring task completion, evidence, and stopping behavior — not fluency.
    The core objective

    By week 16 you can independently diagnose why an Agent made a bad decision — and verify the repair through logs and reruns.

    What you will be able to do

    1. 01Build an Agent that uses tools and knows when to stop.
    2. 02Deliver document research or data analysis with inspectable evidence.
    3. 03Locate decision failures and verify repairs through logs and reruns.
    4. 04Compare an autonomous experiment strategy against a fixed baseline — fairly.
    Sec. 02 · The sixteen weeks Weeks 01–16 · tap a week to expand

    Four phases, sixteen weeks, one Agent you built yourself.

    Phase 01 · Weeks 01–04

    Build an Agent that acts — and stops.

    From LLM APIs and tool calls to a hand-written Agent loop.

    W01 From a request to a working program

    Use Codex as a coding partner. Call an LLM API, inspect its response, and keep credentials outside source code.

    DELIVERABLEA small program that calls a language model and records its output.

    W02 Give the model tools

    Define tools with clear inputs and outputs. Validate requests and return useful errors when a tool fails.

    DELIVERABLEAn assistant that can choose and call two safe tools.

    W03 Build the Agent loop

    Write the decide–act–observe loop yourself. Add a step limit, a stopping rule, and a readable action log.

    DELIVERABLEAn Agent that completes a bounded task and explains why it stopped.

    W04 Know when to ask a person

    Handle missing information, retries, and tasks beyond the Agent's permissions. Add approval before consequential actions.

    DELIVERABLEAn Agent that pauses for help instead of guessing or looping forever.

    Phase 02 · Weeks 05–08

    Give it a real piece of work.

    Choose document research or data analysis. Make every result inspectable.

    W05 Research across documents

    Search a supplied document set, compare claims, and attach sources to the findings.

    DELIVERABLEA research Agent with traceable citations.

    W06 Analyze data with code

    Let the Agent inspect a table, run analysis, read errors, and revise its code. Verify the result with simple checks.

    DELIVERABLEA data-analysis Agent with code and execution records.

    W07 Make the work inspectable

    Choose document research or data analysis. Define success, retain tool records, and specify when a person must take over.

    DELIVERABLEA complete, inspectable first application.

    W08 Application demonstration ASSESSMENT · 30%

    Demonstrate the application on new instructor-provided cases. Submit sources or data provenance, execution logs, and human handoff conditions.

    DELIVERABLEAssessment 1: a working research or analysis Agent.

    Phase 03 · Weeks 09–12

    Make its decisions worth trusting.

    Evaluate failures, control budgets, and build an Autoresearch prototype.

    W09 Evaluate decisions, not fluent answers

    Build a small test set. Score task completion, tool choices, evidence quality, and stopping behavior.

    DELIVERABLEAn evaluation harness and a clear success rubric.

    W10 Diagnose and repair failures

    Read a failed run, locate the decision that went wrong, make a targeted change, and rerun the case.

    DELIVERABLEA before-and-after failure analysis backed by logs.

    W11 Control the experiment budget

    Set limits on tool calls, runtime, retries, and GPU use. Use the supplied YOLO training and evaluation tools without studying their architecture.

    DELIVERABLEA bounded experiment runner with comparable measurements.

    W12 Autoresearch prototype ASSESSMENT · 30%

    Connect the Agent to the experiment tools. Let it propose a permitted setting, run an experiment, read validation results, and choose its next step.

    DELIVERABLEAssessment 2: an Autoresearch Agent with a visible decision trail.

    Phase 04 · Weeks 13–16

    Prove what it can do.

    Compare fairly, reproduce a peer's work, and verify the final system.

    W13 Compare against fixed search

    Run the Agent and a predefined parameter search with the same data, allowed settings, and budget. Compare quality, cost, and reliability.

    DELIVERABLEA fair comparison, including unsuccessful experiments.

    W14 Have a peer rerun your work

    Exchange projects. Reproduce a run, identify missing assumptions, repair a decision problem, and verify the fix.

    DELIVERABLEA peer reproduction record and a verified revision.

    W15 Prepare the final demonstration

    Freeze the evaluation procedure, check records, and rehearse a live run. Explain limitations and when a person should intervene.

    DELIVERABLEA reproducible submission and a concise demonstration.

    W16 Final demonstration and assessment ASSESSMENT · 40%

    Demonstrate the Agent, justify its decisions, and evaluate it with the instructor-held test set. Explain one failure and the evidence for its repair.

    DELIVERABLEAssessment 3: final system, comparison report, and verified repair.

    Sec. 03 · Signature capstone Autoresearch

    Build an Autoresearch Agent.

    From week 11, you work with supplied YOLO experiment tools. It proposes a permitted setting, runs the experiment, reads validation results, and decides what to try next — inside a budget it cannot change.

    Provided — fixed, cannot be changedThe environment your Agent works in

    • Fixed image dataset and data splits
    • Working YOLO training and evaluation tools
    • Allowed parameter settings and a fixed compute budget
    • A fixed-search baseline; the final test set is held by the instructor

    You build — the actual AgentYour design, your evidence

    • An Agent loop that selects its next experiment from evidence
    • State, logs, stopping rules, and failure recovery
    • Equal-budget comparison and reproducible results

    The loop your Agent runs

    Evidence from logs Propose setting Run · budget-capped Read validation Next · or stop

    Parallel baseline — fixed search: same data · same allowed settings · same budget

    Both are compared on quality · cost · reliability — the final test set is held by the instructor

    What counts as success.

    A defensible, reproducible process: an honest equal-budget comparison against fixed search, a peer who can rerun your work, and one diagnosed failure with a verified repair. A higher YOLO score is not required.

    Sec. 04 · Assessment 30 + 30 + 40

    Three checkpoints. Every one is a demonstration.

    Week 08

    30%

    Application demonstration

    Demonstrate your research or data-analysis Agent on new instructor-provided cases.

    • EVIDENCE
    • Sources or data provenance
    • Execution logs
    • Human handoff conditions

    Week 12

    30%

    Autoresearch prototype

    A working research loop on the supplied experiment tools.

    • EVIDENCE
    • Validation results
    • Budget limits
    • Failure handling

    Week 16

    40%

    Final demonstration and assessment

    The finished system, justified decisions, and the instructor-held test set.

    • EVIDENCE
    • Equal-budget comparison
    • Peer rerun
    • Held-out evaluation
    • One verified decision repair
    Sec. 05 · Fit, requirements & FAQ Questions welcome — email anytime

    Is this course for you?

    A good fit if…

    • YESYou work with documents, data, or processes where results must be checked before anyone acts on them.
    • YESYou can write basic programs — and want the model to do more than chat.
    • YESYou would rather verify a repair through logs and reruns than trust a smooth demo.

    How you will learn

    • 01Start building immediately: use an LLM API in week 1 and write your own Agent loop by week 3.
    • 02Use ready-to-run YOLO tools to focus on experiment decisions, evidence and budgets.
    • 03Improve a failed decision and demonstrate the repair with logs and a rerun.

    Requirements

    • R1Basic programming knowledge
    • R2A laptop you can bring to class
    • R3An active ChatGPT Plus subscription with working Codex access — ready before week 1

    Frequently asked

    Q1 Do I need machine-learning experience?

    No. The course starts at LLM API calls and tool use. The YOLO experiment tools arrive ready to run — you never train a network from scratch or study its internals.

    Q2 What exactly must I have before week 1?

    Basic programming knowledge, a laptop, and an active ChatGPT Plus subscription with working Codex access. That is the whole checklist.

    Q3 How much time does it take each week?

    One 3-hour session for 16 weeks (weekly time: to be announced), plus your own build time — most weeks you are extending your own Agent between sessions.

    Q4 Can I use material from my own work?

    Yes. In weeks 5–8 you choose document research or data analysis, and the best material is usually the documents and data you already understand.

    Q5 What do I walk away with?

    A working research or analysis Agent, an Autoresearch prototype with a visible decision trail, and a repeatable way to diagnose failures and prove repairs.

    Q6 Who teaches it?

    陳勇志. Questions before the semester starts are welcome: chenmagi@cc.ncue.edu.tw.

    Questions? Just write.

    Instructor: 陳勇志 · chenmagi@cc.ncue.edu.tw

    Email chenmagi@cc.ncue.edu.tw

    Fall 2027 · Academic year 2027–28, semester 1
    16 weeks × 3 hours · Weekly time: to be announced