SCOPE & PROVENANCE

A reference, not a rulebook.

Five source collections. Thirty-six approaches and building blocks. One consistent way to understand and use them.

How to read the atlas

Choose the right layer.

A workflow controls execution. A planning pattern shapes a plan. A skill supplies context. An adapter is infrastructure. These are not all competing system architectures.

Match the capability.

The conversation template adapts behavior to one assistant. The runtime template assumes real tools and execution support. Neither creates concurrency, permissions, or durable state by itself.

Check before relying on it.

Prompts, example tasks, practical notes, and diagrams are original editorial material. They are not official vendor prompts or benchmarked guarantees. Test them on your own tasks.

The source library

Reviewed 11 Sep 2026

The catalog normalizes common aliases while preserving important distinctions. The LangGraph page is a companion reference within the LangChain collection. This is not a claim to cover every agentic pattern ever proposed.

01Anthropic

Building effective agents

Workflow vocabulary and the distinction between predefined workflows and agent-controlled execution.

Read the primary reference

10 linked approaches

Direct response Single agent with tools Retrieval-augmented generation Prompt chaining Routing Parallelization Orchestrator–workers Reason–act–observe Evaluator–optimizer Voting & ensembles

02Microsoft

AI agent orchestration patterns

Coordination topologies, handoffs, group conversations, and adaptive task-ledger orchestration.

Read the primary reference

10 linked approaches

Direct response Prompt chaining Parallelization Agent handoffs Group discussion Structured debate Adaptive orchestration Evaluator–optimizer Human-in-the-loop Guardrails

03Google Cloud

Choose a design pattern for your agentic AI system

Single-agent and multi-agent flows, loops, hierarchies, swarms, and human checkpoints.

Read the primary reference

13 linked approaches

Single agent with tools Prompt chaining Routing Parallelization Bounded loop Iterative refinement Custom workflow Orchestrator–workers Hierarchical delegation Swarm collaboration Reason–act–observe Evaluator–optimizer Human-in-the-loop

04CSIRO / Data61

Agent Design Pattern Catalogue

The 18-pattern research catalogue, including goal creation, planning, reflection, and supporting components.

Read the primary reference

18 linked approaches

Retrieval-augmented generation Structured debate Role-based cooperation Goal clarification Proactive goal proposal One-shot planning Incremental planning Single-path plan Multi-path planning Self-reflection Cross-reflection Voting & ensembles Human-in-the-loop Guardrails Agent evaluation Prompt & response adaptation Tool & agent registry Agent adapter

05LangChain

Multi-agent patterns

Subagents, handoffs, skills, routing, and custom workflows.

Read the primary reference

5 linked approaches

Routing Custom workflow Subagents as tools Agent handoffs On-demand skills

06LangChain

Workflows and agents

A companion implementation reference for workflow and agent compositions.

Read the primary reference

4 linked approaches

Prompt chaining Parallelization Orchestrator–workers Evaluator–optimizer

A few names worth separating

Orchestration ≠ subagent invocation

Dynamic decomposition creates work packages. A subagent is a call/return boundary. They often work together, but a specialist call does not require a full planning supervisor.

Self-review ≠ independent verification

A second perspective in one chat can be useful, but it is not a separate agent execution. Cross-reflection separates the reviewer; evaluator–optimizer adds the feedback loop.

One-shot ≠ single-path

One-shot describes planning-call cadence. Single-path describes the number of candidate paths. A one-shot call can generate alternatives; one path can be developed incrementally.

Human reflection ≠ permission to act

Feedback improves a proposal. Explicit approval authorizes a specific action. A system must not substitute one for the other.

These distinctions are editorial synthesis. Follow the linked source definitions and your runtime's actual semantics.

TAKE THE REFERENCE WITH YOU

Yours to adapt.

Export all 72 original system prompts and their example tasks as one Markdown file, with source links.

Privacy: this site makes no AI calls. Theme, prompt mode, and saved patterns use local browser storage. When Cloudflare Zaraz is enabled, the site sends action names and limited details such as pattern IDs and category names to Cloudflare. Search text and copied content are never included in these events. External source links open only when you choose them.