University at Buffalo — The State University of New York
School of Engineering and Applied Sciences
Department of Computer Science
and Engineering

CSE 410 / 510 Introduction to Agentic Engineering

Why consider this course?

Important things to know

If you’re looking for a course that’s flexible, project-focused, and a bit different from the usual offerings, this is a great option.

Course Number:CSE 410 / 510 Special Topics
Instructors:Dr. Joaquin Carbonara & Dr. David Doermann
Semester:Fall 2026
Credits:3 credits
Days & Time:Wednesdays, 5:00–7:40 PM
Location:Clemens 120, North Campus
Instruction Mode:In person
Dates:Aug. 24 – Dec. 7, 2026
Format:Lecture / Lab / Project Studio

Course Description

Introduction to Agentic Engineering is a project-based undergraduate and graduate course on designing, building, testing, and responsibly deploying AI-powered agents: software systems that combine foundation models with tools, data sources, memory, planning, structured workflows, and human oversight. The course treats an agent not as a magic chatbot, but as an engineered system that senses a bounded environment, chooses actions through well-designed interfaces, uses tools, and is evaluated against explicit objectives and safety constraints.

Students will study the full agentic engineering stack: prompt and context design, tool/function calling, retrieval-augmented generation, embeddings and vector search, workflow graphs, state machines, planning, multi-agent coordination, evaluation, observability, security, and human-in-the-loop design. Mathematical topics will be introduced as engineering tools: vector spaces and similarity for embeddings; probability, uncertainty, and calibration; graph search and finite-state workflows; Markov decision processes and reinforcement-learning concepts; optimization and loss functions; and statistical evaluation through metrics, confidence intervals, bootstrapping, and error analysis.

The course culminates in a team project in which students build a useful agent for a bounded real-world task, document its architecture, create an evaluation suite, analyze risks, and present a working prototype.

Prerequisites & Target Audience

Intended audience: This course is offered under two course numbers. CSE 410 is for upper-division undergraduate students in Computer Science, Computer Engineering, Data Science, Cybersecurity, and related technical programs. CSE 510 is the graduate section for MS and PhD students seeking a graduate-level elective in agentic AI engineering.

Prerequisites: CSE 250 Data Structures or equivalent programming maturity, or approval by the instructor. Students should be comfortable writing Python programs, using Git/GitHub, reading technical documentation, and debugging code.

Recommended background (not required): CSE 331 Algorithms and Complexity; CSE 368 Introduction to Artificial Intelligence; CSE 442 Software Engineering; CSE 435 Information Retrieval; CSE 474 Introduction to Machine Learning, or equivalent experience. Mathematical maturity will help, but required mathematical ideas will be introduced in context.

Learning Outcomes

Upon successful completion of this course, students will be able to:

  1. Explain the difference between chatbots, workflows, tool-using LLM applications, and autonomous/semi-autonomous agents.
  2. Design a bounded agentic system by specifying its environment, goals, tools, memory, state, inputs, outputs, and failure modes.
  3. Implement agentic workflows using Python and contemporary frameworks or lightweight custom code.
  4. Use tool/function calling, structured outputs, retrieval, and embeddings to connect models to external information and actions.
  5. Model agent workflows using graphs, state machines, and planning/search concepts.
  6. Apply mathematical tools from linear algebra, probability, graph theory, optimization, reinforcement learning, and statistics to agent design and evaluation.
  7. Build evaluation harnesses for agent behavior, including task-level metrics, error taxonomies, regression tests, and human review.
  8. Analyze ethical, privacy, security, and social risks in agentic systems.
  9. Communicate technical designs through architecture diagrams, documentation, demos, and reflective analysis.

Mathematical Themes

The course introduces mathematical ideas in the context of engineering problems, not as prerequisites. Topics include:

Linear Algebra for EmbeddingsVectors, inner products, norms, cosine similarity, dimensionality, nearest-neighbor search.
Information RetrievalTF-IDF, BM25 intuition, dense retrieval, hybrid retrieval, reranking.
Probability & UncertaintyConditional probability, Bayes' rule, likelihood, calibration, confidence, expected loss.
Graphs & StateDirected graphs, workflow graphs, finite-state machines, graph search, planning, dependency structures.
OptimizationObjectives, loss functions, gradient descent intuition, prompt optimization, preference optimization.
Sequential Decision-MakingStates, actions, rewards, policies, Markov decision processes, RL vocabulary.
Evaluation StatisticsPrecision, recall, F1, task success rate, confusion matrices, bootstrapping, confidence intervals, paired comparisons.
Systems TradeoffsLatency, cost, throughput, error propagation, retry policies, caching, reliability.

Tentative Fall 2026 Schedule

UB's standard Fall 2026 session runs August 24 – December 7, 2026. The schedule below assumes a standard 15-week semester and will be adjusted once official meeting days are confirmed.

Weeks 1–8

WkDatesTopicsMath / Technical FocusMajor Work
1Aug. 24–28 What is agentic engineering? Agents, environments, tools, goals, constraints. Agent = environment + actions + policy; system boundaries Reading note 1
2Aug. 31–Sep. 4 Foundation models as components in software systems. Prompting, context, structured outputs. Probability over tokens; uncertainty; structured outputs Lab 1 assigned
3Sep. 8–11 Tool/function calling and the ReAct pattern. Reasoning/action loops; action spaces; error propagation Lab 1 due
4Sep. 14–18 Retrieval-augmented generation and context construction. Vectors, cosine similarity, TF-IDF / BM25 intuition Lab 2 assigned
5Sep. 21–25 Embeddings, vector search, reranking, and memory. Nearest-neighbor search; precision/recall; retrieval evaluation Math mini-assignment 1
6Sep. 28–Oct. 2 Agent workflows: graphs, state machines, LangGraph/custom orchestration. Directed graphs, finite-state machines, DAGs Lab 2 due; Lab 3 assigned
7Oct. 5–9 Planning and decomposition. When should an agent plan? Search, heuristics, task decomposition, cost functions Project proposal due
8Oct. 14–16 Human-agent interaction, interface design, and human-in-the-loop. Decision thresholds; expected cost of errors Midterm architecture review assigned

Weeks 9–Final

WkDatesTopicsMath / Technical FocusMajor Work
9Oct. 19–23 Evaluation of agents. From single prompts to task-level evals. Metrics, benchmarks, confidence intervals, bootstrapping Lab 3 due; Lab 4 assigned
10Oct. 26–30 Security, privacy, prompt injection, tool misuse, and sandboxing. Threat models; invariants; permission boundaries Midterm architecture review due
11Nov. 2–6 Multi-agent systems and coordination. Graph coordination, communication protocols, game-theoretic intuition Lab 4 due; Lab 5 assigned
12Nov. 9–13 Coding agents, software engineering agents, testing, CI, and observability. Regression tests, traces, reliability statistics Project checkpoint
13Nov. 16–20 Agents for data science, scientific discovery, and education. Experimental design; causal caution; model limitations Lab 5 due
14Nov. 23–24 Responsible deployment and final project studio. Risk analysis, calibration, accountability Draft project report due
15Nov. 30–Dec. 7 Final demos, critique, and synthesis. Evidence-based evaluation of agent usefulness Final demos in class
FinalDec. 9–16 Final project submission at assigned exam time. Final evaluation report Final report & repository due

Major Assignments

Lab 1 A Minimal Tool-Using Agent

Build a small Python agent that can choose among a calculator, a file reader, and a simple search or lookup function. Submit code, a short design note, and examples of successful and failed runs.

Lab 2 Retrieval and Memory

Create a small RAG pipeline using a course-selected document collection. Compare keyword retrieval, dense retrieval, and/or hybrid retrieval. Include cosine similarity calculations and an error analysis.

Lab 3 Workflow Graphs and State Machines

Implement an agentic workflow as a graph or state machine. The system should include explicit states, transitions, failure handling, and logging.

Lab 4 Evaluation Harness

Design an evaluation suite for a bounded agent task. Include test cases, success criteria, metrics, a failure taxonomy, and a reproducible evaluation script.

Lab 5 Human-in-the-Loop and Safety Boundaries

Add a human approval step, structured outputs, guardrails, or permission boundaries to an agentic workflow. Analyze where autonomy is helpful and where it should be restricted.

Final Team Project: Useful Bounded Agent

Teams will design and build a working agent for a clearly bounded task. Deliverables include a problem statement, environment and tool inventory, architecture diagram, implementation repository, evaluation dataset or script, results and error analysis, security/privacy/ethical risk assessment, demo, and final written report.

Grading

ComponentWeight
Weekly reading / design notes10%
Labs and technical exercises25%
Mathematical mini-assignments15%
Midterm architecture review15%
Final team project25%
Participation, peer review, and professionalism10%
Total100%

Required and Recommended Materials

There is no single required textbook. Readings will be a mix of book chapters, technical documentation, research papers, and practitioner essays.

Tools and Technical Environment

Students will use Python, Git/GitHub, Jupyter notebooks, command-line tools, and one or more agent frameworks. The course may use OpenAI-compatible APIs, open-source/local models, LangChain/LangGraph, LlamaIndex, vector databases or local vector stores, and lightweight web or API frameworks. Exact tools may be adjusted based on UB infrastructure, student access, and current best practices in Fall 2026.

Students should expect to document their work clearly and maintain reproducible repositories. Where paid APIs are used, alternatives or credits will be provided when possible. Students may also use local models through tools such as Ollama when technically appropriate.

Course Policies

Attendance and Participation
Because the course is project-based and involves labs, design reviews, and peer critique, regular attendance is expected. Students are responsible for material, announcements, and deadlines discussed in class.
AI Tool Use Policy
Students may use AI tools as part of the course because the course is about building and understanding AI systems. However, students must use them transparently and professionally. All submitted work must include an AI Use Statement explaining which tools were used, for what purpose, and what parts of the work were substantially generated, revised, debugged, or evaluated with AI assistance. Students remain responsible for correctness, originality, citations, code quality, and ethical use.
Collaboration
Collaboration is encouraged for discussion, debugging, and project work, but individual submissions must reflect each student's own understanding unless the assignment is explicitly designated as a team assignment. Team projects must include a contribution statement.
Reproducibility
Assignments involving code must include enough information to reproduce results: dependencies, instructions, test cases, example inputs, expected outputs, and any limitations. Students should use Git regularly and submit clean repositories.
Accessibility and Inclusion
The course will emphasize clear documentation, readable code, and accessible communication. Students should contact the instructor and UB Accessibility Resources as early as possible if they need accommodations.
Academic Integrity
Students are expected to follow UB academic integrity policies. In this course, academic integrity includes honest reporting of AI assistance, data sources, code sources, evaluation results, and project limitations.