Core Single-Agent Patterns
The ReAct pattern (Reasoning and Acting) structures agent behavior as an alternating cycle: the agent produces reasoning about its current situation, takes an action, observes the result, and reasons again before taking the next action. This interleaving of reasoning and action makes the agent's behavior more interpretable and correctable than approaches where the agent acts without explicit intermediate reasoning steps. The plan-and-execute pattern separates planning from execution: a planner generates a full sequence of steps, and an executor carries them out. This suits long-horizon tasks where maintaining a coherent overall plan matters more than adapting to each intermediate result. The reflection pattern adds a self-evaluation step: after producing an output or completing a step, the agent critiques its own work and revises before proceeding, which improves output quality at the cost of additional compute and latency.
Multi-Agent Patterns
When tasks benefit from parallelism or require domain-specialized handling, multi-agent patterns distribute work across agent instances. The orchestrator-worker pattern uses a coordinating agent to break a task into subtasks and delegate each to a worker agent suited to that type of work. The debate or multi-perspective pattern runs multiple agents on the same task and uses agreement or disagreement among them to improve output quality or identify uncertainty. Supervisor patterns allow a monitoring agent to review worker outputs and intervene when quality falls below a threshold. The appropriate pattern depends on how the task decomposes, how independently subtasks can be run, and what the failure modes of each agent are. Simpler tasks often do not benefit from multi-agent patterns, and the coordination overhead can reduce overall reliability.