Knowledge Vault

Agent Patterns

2 min read

Overview

Agent patterns describe architectural approaches for building autonomous AI systems that can perceive their environment, make decisions, and take actions to achieve goals. Unlike simple prompt-response interactions, agents operate in loops — observing, reasoning, acting, and incorporating feedback. Common patterns include ReAct (reasoning + acting), plan-and-execute (decompose then delegate), and multi-agent systems where specialized agents collaborate. The key design challenge is balancing autonomy with reliability: agents must be capable enough to handle complex tasks but constrained enough to avoid harmful or wasteful actions.

Key Concepts

  • ReAct loop — The agent alternates between reasoning (thinking about what to do) and acting (calling tools or producing output). Each observation feeds the next reasoning step.
  • Plan-and-execute — Agent first creates a plan, then executes each step. Better for multi-step tasks where upfront decomposition improves accuracy.
  • Multi-agent systems — Multiple specialized agents collaborate on a task. Patterns include orchestrator-worker, debate, and ensemble voting.
  • Tool selection — Agents choose which tools to invoke based on the current task. Requires the model to understand tool capabilities and constraints.
  • Memory — Short-term (context window), medium-term (conversation summaries), and long-term (persistent storage) memory enable agents to maintain state across interactions.
  • Guardrails — Constraints that prevent agents from taking harmful, expensive, or unauthorized actions. Include output validation, confirmation prompts, and scope limitations.

Trade-offs

  • More autonomy increases capability but also risk of errors compounding
  • Multi-agent systems add coordination overhead and debugging complexity
  • Planning upfront is slower but more reliable for complex tasks
  • Guardrails reduce risk but also reduce agent flexibility

Graph View

Backlinks