AI Agents Bootcamp

Schedule

The time frames are only estimates and may vary according to how the class is progressing.

DAY 1: Foundations of Agent Design and Multi-Agent Collaboration

Foundations of agent design and task decomposition (60 minutes)

• Presentation: State machines, structured state transitions, and how to translate real-world tasks into agent workflows using LangGraph
• Demo: Transforming a stateless LLM into a stateful agent with basic tool use
• Hands-on exercise: Decompose a given small task and implement it as a LangGraph workflow with state and one tool
• Q&A
• Break

Making the agent smarter (60 minutes)

• Presentation: Adding intelligence and optimization at test time using structured reasoning patterns such as chain of thought and ReAct
• Demo: Implementing chain of thought and ReAct for an agent from scratch
• Hands-on exercise: Compare a simple single agent to a planner; judge setup and observe how structured reasoning and test-time compute affect the quality of the outcome
• Q&A

Adding human oversight to your system (60 minutes)

• Presentation: How much autonomy agents have in practice; implementing safeguards with human-in-the-loop patterns
• Demo: Adding human-in-the-loop checkpoints in LangGraph to approve or reject actions, edit graph state, and request additional human input before sensitive steps
• Hands-on exercise: Insert a human-in-the-loop gate into your own agent workflow
• Q&A
• Break

From single agents to multi-agent systems (60 minutes)

• Presentation: Scaling from a single, tool-using agent to a multi-agent system using collaboration patterns; when to apply each
• Demo: Building a supervisor-based multi-agent system in LangGraph where a coordinator agent delegates tasks to specialist agents
• Hands-on exercise: Turn your existing agent into a two-agent workflow and test how responsibilities and handoffs change system behavior
• Q&A
• Break

Communication patterns in multi-agent systems (60 minutes)

• Presentation: How multi-agent systems coordinate through message passing, shared memory access patterns, and event-driven communication; when to choose each method; how communication shapes collaboration patterns, latency, handoff reliability, and system throughput
• Demo: How different multi-agent collaboration patterns work together with communication patterns; which ones to combine for which use case
• Hands-on exercise: Exchange a communication pattern in a predefined MAS and observe how it affects agent coordination and communication
• Q&A

DAY 2: Production-Ready Multi-Agent Systems

From unstructured and multimodal inputs to structured agent-ready data (60 minutes)

• Presentation: Optimizing agents for structured, unstructured, and multimodal data; aligning their outputs through schema-based prompting with Pydantic validation; enabling stable multi-agent communication using the Model Context Protocol (MCP)
• Demo: Transforming unstructured data retrieved from an MCP server into agent-digestible structured data
• Hands-on exercise: Take a free-form task description and define a minimal schema, configuration, and prompt so that an agent can consistently turn it into structured data for a downstream tool or coding workflow
• Q&A
• Break

Memory design and shared context strategies (60 minutes)

• Presentation: How episodic memory preserves task-level context; how procedural memory enables routines that can be reused across tasks; capturing higher-level reasoning patterns through long-term memory structures; determining what information should be stored, when memory should be updated, and how retrieval shapes subsequent decisions; checkpointing as a mechanism for preserving state
• Demo: A LangGraph workflow that maintains episodic and procedural memory and uses checkpoints; how memory is read and written during execution; how time travel allows stepping back to an earlier state to inspect the reasoning path
• Hands-on exercise: Extend your agent by adding a simple episodic memory that stores each tool result and a procedural memory entry that records a small reusable routine
• Q&A
• Break

Choosing the right models and architectural patterns for agent reasoning (60 minutes)

• Presentation: How architectural decisions shape the capability, speed, and stability of an agentic system; comparing planner executor setups, supervisor based coordination, and unified agent designs; assigning models to roles based on their strengths; when to use thinking or no thinking inference modes; KV cache behavior; context engineering, and the need for stable prefixes when designing long running workflows
• Demo: Comparing two LangGraph architectures, one using a planner with a thinking mode enabled and one using a unified agent with no thinking mode
• Hands-on exercise: Modify a predefined workflow by switching the planning component between thinking and no thinking modes and observe the change in latency and decision quality; run the same workflow with a dense model and a smaller MoE model to compare behavior
• Q&A
• Break

Observability, monitoring, and benchmarking (60 minutes)

• Presentation: Observing, measuring, and evaluating agent behavior in production; building task-specific benchmarks for multi-agent systems
• Demo: Monitoring and evaluating an agent workflow with monitoring hooks, logging, and a custom evaluation loop
• Hands-on exercise: Add tracking to your workflow with one simple logging function that prints each state transition; run the workflow twice with small changes in the input and compare the differences in the logged steps to evaluate stability
• Q&A
• Break

Security, guardrails, and secure execution environments (50 minutes)

• Presentation: The agentic threat landscape (reasoning manipulation, memory poisoning, tool misuse, communication corruption), safeguarding coding agents
• Demo: Securing an agentic workflow using OmegaConf-driven security policies, Pydantic schema validation, MCP tool governance, and PRP-based coding constraints
• Group discussion: Review a small predefined scenario involving an unsafe or ambiguous agent action and discuss where the vulnerabilities lie, what should be validated, and what mitigation strategy is most appropriate

Wrap-up and Q&A (10 minutes)