15 AI Agent Design Patterns, Distilled From Building 400+ Agents
MaryamMiradi · x · 2026-10-19
AI engineer Maryam Miradi shares 15 design patterns for AI agents, organized in three tiers based on her experience building 400+ agents:
Single-agent patterns: ReAct (reason before each tool call), Plan-and-Execute, Reflection/Self-Critique, and Tool Use/Function Calling.
Multi-agent orchestration: Orchestrator-Subagent, Supervisor with quality gates, Hierarchical Agents, Sequential Pipeline, Parallel Fan-Out/Fan-In, MapReduce, and Debate/Adversarial (agents argue, a judge decides).
Iterative & feedback loops: Evaluator-Optimizer, Critic-Actor, Self-Healing/Retry Loop, and Human-in-the-Loop checkpoints.
Her framing: single-agent patterns define what one agent can do alone, multi-agent patterns define coordination under load, and feedback loops define how systems recover and improve.
More from coding & agent
- AI agents cost 20x more than conventional software, warns architect's enterprise breakdown — DavidLinthicum · 2026-10-19
- Indie dev shares UI iteration workflow: three mock options for every visible change — SeanOliver · 2026-10-19
- Reddit dev splits Claude Code tasks by complexity to cut token burn, but sub-agents backfire — alatmaca · 2026-10-19
- Opus 5.5 builds a realistic 3D ocean exploration scene in 4,100 lines of code — techartist_ · 2026-10-19
- DeepSeek V4.1 gets smarter but discards its best ideas; Opus struggles to rescue the mess — teortaxesTex · 2026-10-19
- eduframe-mcp open-sources natural language lead management for Eduframe via MCP — modelcontextprotocol · 2026-10-19