List of Agentic AI Patterns and Architectures. This is still an evolving field, so the list is updated as of 04-Jan-2026.
- Single-Agent and Workflow Patterns
- ReAct (Reason + Act) loop – he classic dynamic loop: Thought → Action → Observation. It allows the agent to “figure things out” step-by-step but can be unstable in long runs.
- Prompt‑chaining / sequential pipeline – Deterministic flows where the output of step A always becomes the input of step B. Used for reliable tasks like “Research → Summarize → Translate”
- Planner–executor (plan‑then‑act) – Separates “Thinking” from “Doing.” The agent first writes a manifest of 5 steps, then executes them deterministically. Reduces the chance of the agent getting “distracted” mid-task.
- Self‑reflection / self‑critique loop
- Tool‑augmented single agent (Function Calling) – An LLM that detects when to call an API (e.g., OpenAI Function Calling) to get data before answering.
- Router / dispatcher agents – A gateway that classifies user intent (e.g., “Billing” vs. “Tech Support”) and routes the query to a specialized downstream agent or prompt.
- Evaluator-Optimizer (Reflexion) – A “quality assurance” pattern. One LLM call generates a draft, and a second call acts as a “Critic” to score it and request improvements. Proven to significantly boost coding/writing quality.
- Multi-Agent Patterns
- Supervisor / orchestrator + worker agents / Hub-and-Spoke – The Multi-Agent Standard. A central “Manager” agent breaks down the task, delegates sub-tasks to specific “Worker” agents (Coder, Researcher), and aggregates their answers
- Parallel specialist agents (fan‑out/fan‑in) – Map-Reduce for LLMs. A manager spins up multiple identical agents to process different files or tasks simultaneously, then aggregates the results.
- Hand-offs / Swarm Pattern – Decentralized control. Agent A realizes it can’t finish a task and “transfers” the user directly to Agent B (e.g., Triage Agent → Sales Agent), with no central manager involved.
- Role‑based crew / team pattern – Agents simulate a human team meeting, often conversing with each other in a round-robin or hierarchical fashion to solve a problem.
- Agents‑as‑Tools Pattern – Encapsulation. A complex agent (e.g., a “SQL Writer Agent”) is wrapped up and given as a single “tool” to a higher-level agent.
- Production Reliability & Safety Patterns
- Human-in-the-Loop (HITL) The agent pauses execution at critical junctures—like sending an email or spending budget—to wait for human approval via API/UI.
- Compliance / Guardrail Co-pilot – A secondary, lower-intelligence “Monitor” model watches the main agent’s output in real-time. It blocks toxic content, PII leaks, or off-topic actions before they reach the user.
- Corrective & Self‑healing Agents – Agents equipped with error-handling logic. If a tool fails (e.g., API timeout), the agent analyzes the error message and tries a different parameter or alternative tool.
- Advanced / Frontier Architectures
- Memory‑centric / persona‑persistent agents – Agents with long-term memory layers (like MemGPT) that remember user details across sessions weeks apart
- Environment‑/simulation‑based agents – Agents that “live” inside a sandbox (like a coding environment or game) and learn via Reinforcement Learning or trial-and-error.
- Hierarchical Multi‑level Planning – Extremely complex systems with multiple layers of supervisors (e.g., a CEO Agent managing Manager Agents managing Worker Agents).
