AI agents are defined by their autonomy and reasoning capabilities, distinguishing them from reactive systems like generative AI, RAG, and traditional chatbots.
An agent is defined in code with a model, description, tools, and a set of natural language instructions that shape its core behavior and goals.
Breaking down complex tasks into smaller, specialized sub-agents that collaborate improves modularity and reusability in a system.
Graph-based workflows offer more predictable execution paths than LLM-driven routing, with pre-built patterns like sequential, loop, and parallel agents.
Input and output schemas enforce data consistency for agent interactions, while skills help manage the context window by loading functionality on demand.
An agent's architecture consists of tools, a model, orchestration, and a runtime, with emerging protocols like MCP simplifying tool integration and discovery.
Evaluating non-deterministic agents requires assessing both the final output and the trajectory of steps taken, before deploying them to platforms like Cloud Run or GKE.
Building reliable agents involves addressing challenges like performance and cost, while implementing guardrails and considering ethical issues like explainability and fairness.
Jobs that call for the skills explored in this talk.