An orchestrator decomposes high-level goals into sub-tasks, routes them to specialized worker agents, aggregates results, and maintains a shared context window across the entire execution graph.
ReAct, CoT, and tree-of-thought planning systems decompose goals into action sequences. The planner emits thought → action → observation loops, with each step grounded by tool calls and memory lookups.
Agents select and invoke tools — code interpreters, web search, file systems, APIs, databases — via structured function calling. The tool router resolves schema, validates arguments, executes, and pipes results back into context.
Agents maintain multiple memory tiers: in-context ephemeral working memory, episodic long-term memory (vector store), semantic procedural memory (knowledge graphs), and external persistent storage.
Societies of specialized agents — researchers, coders, critics, coordinators — communicate via structured message passing. Emergence arises from local interaction rules, role specialization, and shared memory access.
Agent performance is measured across task completion rate, tool use efficiency, context faithfulness, hallucination rate, and multi-step reasoning depth. Evals drive iteration cycles across all agent system components.
The researchers, engineers, and organizations shaping autonomous AI agent architectures.
Core vocabulary for autonomous agent system design and deployment.