Project 1 proposal
Project selection, scope, architecture, evaluation plan, and milestones.
INF 385T · Fall 2026 · Graduate Course
Language Models
and Agentic Workflows
Design reliable systems that retrieve context, call tools, maintain state, adapt through bounded loops, and remain observable, evaluable, and accountable.
Course throughline
Build autonomy deliberately.
The course begins with model interfaces and context, then layers in deterministic orchestration, adaptive control, evaluation, planning, interoperability, multi-agent coordination, human oversight, security, and production reliability. Each week pairs concepts with hands-on implementation.
Fall 2026
Key milestones
Major checkpoints from the current course syllabus.
Paper-based quiz
Individual assessment on workflows, agents, context, tools, loops, and evaluation.
Project 1 completion
Working system, evaluation artifacts, source code, and technical report.
Project 1 presentations
Full presentation and working demonstration with architecture and evaluation defense.
Project 2 presentations
Production-oriented agentic system presentations and live demonstrations.
Project 2 report
Final technical report, reproducibility materials, and evaluation analysis.
Weekly course map
Sixteen modules, one cumulative system
Slides and notebooks are available below for Weeks 1–3. Additional materials will be added as the course progresses.
01Materials ready
From Language Models to Agentic Systems
Models, workflows, agents, autonomy, system components, and when an agent is—or is not—the right architecture.
Lab: Compare unstructured and structured model responses; establish reproducible experiments.
02Materials ready
Structured Model Interfaces and Tool Use
Schemas, structured outputs, function calling, tool contracts, and model–tool failure modes.
Lab: Validate schema-constrained output and diagnose incorrect tool selection and arguments.
03Materials ready
Context Engineering: Retrieval, Memory, and State
Context selection, dense and sparse retrieval, working and persistent memory, provenance, and grounding.
Lab: Build a retrieval pipeline, add memory, inspect evidence, and evaluate grounding failures.
04
Deterministic Agentic Workflows
Prompt chains, routing, parallelization, evaluator–optimizer patterns, state machines, and task graphs.
Lab: Build a stateful workflow with branches, validation gates, retries, and execution visualization.
05
Single-Agent Loops and Adaptive Control
Plan–act–observe loops, dynamic tool choice, stopping conditions, bounded autonomy, and recovery.
Lab: Implement a multi-tool loop with iteration, cost, failure, and completion limits.
06
Agent Evaluation, Tracing, and Reliability
Trajectory evaluation, grounding, tool-call correctness, failure taxonomies, graders, and regression tests.
Lab: Instrument traces, construct an evaluation set, and compare workflow variants.
07
Planning, Reflection, and Verification
Task decomposition, plan execution, replanning, critics, verifiers, self-critique limits, and recovery.
Lab: Add planning and verification, then measure quality, latency, and cost trade-offs.
08Quiz + lab
Paper-Based Quiz and Agent Interoperability
Individual quiz followed by MCP concepts, capability discovery, clients, servers, resources, and tools.
Lab: Connect an agent to an interoperable tool service.
09
Multi-Agent Systems and Project 1 Workshop
Supervisors, specialists, handoffs, delegation, shared state, communication, and coordination failures.
Workshop: Compare single- and multi-agent designs and test Project 1 workflows.
10
Human–Agent Interaction
Approval boundaries, escalation, autonomy levels, transparency, interruptibility, correction, and trust.
Lab: Add approval gates, pause/resume, cancellation, and risk-based escalation.
11Presentations
Group Project 1 Presentations
Architecture explanation, working demonstration, evaluation findings, failure analysis, and critique.
Deliverables: System, code, evaluation artifacts, presentation, demo, and technical report.
12
Agent Security, Safety, and Governance
Prompt injection, excessive agency, least privilege, sandboxing, authorization, red teaming, and governance.
Lab: Conduct adversarial tests and implement tool, input, output, and action guardrails.
13
Industry Keynote, Deployment, and Project Workshop
Production challenges, durable execution, checkpoints, recovery, idempotency, monitoring, latency, and cost.
Workshop: Stress-test Project 2 and review operational evidence.
14No class
Fall Break / Thanksgiving
No class meeting. Use the break to consolidate Project 2 implementation and evaluation evidence.
15Presentations
Group Project 2 Final Presentations
Live demonstrations and defense of architecture, evaluation, robustness, security, and human oversight.
16
Offline Report Preparation
Incorporate feedback, complete technical documentation, finalize evaluation analysis, and submit reproducibility materials.
Final report due: December 10, 2026.
Course objective