Architecture
How agents reason, coordinate and connect.
Plan-and-Execute The agent generates a complete plan upfront and executes it sequentially. Predictable, auditable, rigid. ReAct (Reasoning + Acting) The agent interleaves thinking and doing in a tight loop. The plan emerges from execution, not before it. Coordination patterns The five foundational patterns for composing LLM calls and agents, from prompt chaining to evaluator-optimizer. Multi-agent topologies How agents are wired together: hub-and-spoke, hierarchical, flat mesh and swarm, and why production picks the hub. The interoperability stack How agents and tools connect across boundaries: WebMCP, MCP and A2A, converging under the Linux Foundation. MCP (Model Context Protocol) The de facto agent-to-tool protocol. Created by Anthropic, now at the Linux Foundation. A2A (Agent-to-Agent) The protocol that gives agents colleagues. Discovery, communication and delegation across boundaries. Foundation model Large-scale, broad-data, many tasks. Characterized by emergence, adaptation and homogenization. Necessary but not an agent. Tool use / function calling Function calling is the mechanism. Tool use is the capability it enables. The model never executes anything itself. Agent vs Agentic An AI agent is the discrete software entity. Agentic AI is the paradigm, not the entity. Agent vs Workflow A workflow executes a predefined path. An agent directs its own process, in a loop, on feedback from the environment.