Agentic Design Patterns
A field-aware reference covering 23 patterns for building agentic systems — organized by the question each pattern answers, not by the year it was named.
Press / to focus. Searches name, alternative names, summary, and layer.
23 patterns found
Layer 1 — Topology / Control Flow
How should my agent calls be arranged and sequenced?
Patterns that govern the structural shape of an agent system — how steps are ordered, branched, parallelized, and delegated across single-agent and multi-agent topologies.
Single-agent (10)
Prompt Chaining
A fixed sequence of LLM calls, each consuming the previous step's output.
Open pattern →
Routing
Classify the input, then dispatch it to the handler that fits.
Open pattern →
Parallelization
Fan a fixed set of LLM calls out at once, collapse with a deterministic aggregator.
Open pattern →
Planning
Agent drafts a multi-step plan, executes it, and rewrites the tail when reality bites.
Open pattern →
Tool Use / ReAct
Agent interleaves a thought, a tool call, and the result until it can answer.
Open pattern →
Code Agent
Tool-using agent whose primary toolkit is a codebase, an editor, and a test runner.
Open pattern →
Evaluator-Optimizer
Generate a draft, score it against a rubric, refine until the critic stops complaining.
Open pattern →
RAG
Retrieve passages from an external store and condition the model on them.
Open pattern →
Agentic RAG
Agent reads partial passages, decides whether to re-query, and stops on its own.
Open pattern →
Reflexion
Agent writes self-critiques into memory to improve next attempts.
Open pattern →
Multi-agent (3)
Orchestrator-Workers
A central planner spawns worker agents per task, then merges their outputs.
Open pattern →
Multi-Agent Debate
Several agents argue, critique each other, and converge on a single answer.
Open pattern →
Handoffs / Swarm
One agent owns the conversation; transfer-as-tool-call hands it to a specialist.
Open pattern →
Layer 2 — Quality & Control Gates
How do I catch errors, enforce safety, and know when to stop?
Patterns for verifying outputs, enforcing policy constraints, and injecting human judgment at the points where autonomous decisions carry meaningful risk.
Guardrails
Layered checks around the model that block unsafe input and output before either ships.
Open pattern →
Human in the Loop
Agent pauses on a designated step and waits for a human signal before continuing.
Open pattern →
Evaluation (LLM-as-Judge)
Score model outputs with another LLM applying a written rubric.
Open pattern →
Layer 3 — State & Context
How does my agent retain and manage what it knows across steps?
Patterns that manage working memory, persistent state, and context compression — how an agent carries information across turns, steps, and sessions.
Memory Management
Tier the agent into working, episodic, and semantic stores it can read and edit.
Open pattern →
Context Engineering
Pick which tokens enter the window, in what order, with what weighting.
Open pattern →
Checkpointing
Agent persists its run state at safe boundaries so a crash mid-loop loses no work.
Open pattern →
Layer 4 — Interfaces & Transport
How do agents and tools communicate with each other and external systems?
Patterns governing the protocols, wire formats, and integration surfaces by which agents expose capabilities to callers and consume capabilities from external systems.
MCP
Open protocol for agents to discover and call tools across vendors over one wire format.
Open pattern →
A2A
Cross-runtime agents discover each other by URL and exchange tasks over HTTP+JSON.
Open pattern →
Streaming
Deliver partial output as it is generated, not after the model has finished.
Open pattern →
Layer 5 — Methodology
What principles should guide how I design and operate agent systems?
Cross-cutting stances and design philosophies that shape how you approach agentic system design as a whole — beyond any single component or interaction pattern.