Detailed Course Outline
1. Fundamentals of Agent Abstraction and LLMs
- Discuss LLM capabilities & pitfalls
- Introduce agents as a task decomposition abstraction.
- Demonstrate minimal agent with free-text LLM calls.
2. Structured Output & Basic Fulfillment Mechanisms
- Bottlenecking LLMs with JSON/task-based outputs.
- Ensure domain alignment & stable schema enforcement.
- Introduction to cognitive architectures.
3. Retrieval Mechanisms & Environmental Tooling
- Formalize environment access strategies for agents to interface with other systems.
- Develop tool interfaces for external data repositories (DBs, APIs)
- Use vector-RAG-coded for semantic retrieval over document sets.
4. Knowledge Graphs & Document Graphs
- Plan progression of data from raw docs to canonical forms.
- Motivate threshold/equilibrium objectives for driving event loop.
- Build state pools/ontologies for robust domain coverage
5. Multi-Agent Systems & Frameworks
- Decompose tasks among specialized agents
- Formalize communication buffers and process distribution schemes.
- Differentiate between different frameworks and their unique approaches.
6. Data Flywheels & System Hardening
- Capture usage logs, refining domain constraints, or sub-models
- Implement human-in-the-loop oversight for error correction
- Iterative improvement & pipeline simplification using real/synthetic data.
7. Scaling & Productionalization
- Discuss production-oriented considerations like resource management, concurrency, resource utilization, multi-tenancy
- Motivate framework-agnostic modular deployments (meta-frameworks) and their selection criteria.
8. Final Assessment
- Deploy an agent endpoint that can support multiple different interactions.
- Run a distributed dialog loop across the deployed server to assess satisfaction.
9.1. [Optional] Real-Time Agents
- Discuss multimodal considerations and agentic use-cases that interact with the physical world.
- Explore recent advances in robotics, audio systems, and world models.
9.2. [Optional] Responsible Agents
- Discuss common failure modes in software design that introduce unfairness, liability, and poor software experiences.
- Consider checks-and-balances systems, standards creation, and evaluation needs.