Agent Design Patterns
A systematic catalogue of 18 architectural patterns for building foundation model-based agents — from goal creation and planning to multi-agent cooperation and safety guardrails.
12 lessons
01Foundation Model Agents & the Pattern Catalogue
The anatomy of an FM-based agent, the five challenge dimensions that make agent design hard, and the six pattern groups that address them.
Goal Creation Patterns
How agents capture user intent: the Passive Goal Creator waits for explicit requests, the Proactive Goal Creator monitors context and acts without being asked.
Context & Knowledge Patterns
The Prompt/Response Optimiser refines how the agent talks to the model; RAG augments what the model knows at query time.
Planning Patterns
Single-path Plan Generator commits to one ordered action sequence; Multi-path generates alternatives and selects the best — the right choice depends on how costly a wrong plan would be.
Model Querying Patterns
One-shot vs incremental model querying — how agents structure their interactions with foundation models to balance cost, latency, and reasoning quality.
Reflection Patterns
Self-Reflection, Cross-Reflection, and Human Reflection — how agents structure critique and revision loops to reduce hallucinations and improve output quality.
Multi-Agent Cooperation Patterns
Voting-based, Role-based, and Debate-based cooperation — how multiple agents divide work, combine perspectives, and produce higher-quality outputs than any single agent.
Safety, Registry & Adaptation Patterns
Multimodal Guardrails, Tool/Agent Registry, Agent Adapter, and Explainability — the infrastructure patterns that make agents safe, extensible, and maintainable in production.
Tool Use, Function Calling & MCP
The engineering substrate beneath every agent: how a model emits a structured tool call, how your runtime executes it, why errors must be handled, and how the Model Context Protocol standardises the whole integration.
Evaluating Agents
Why agent evaluation is harder than model evaluation: outcome vs trajectory metrics, LLM-as-judge on traces, agent benchmarks, and the cost/latency budgets that decide whether an agent ships.
Deploying Agents in Production
From a working loop to a reliable service: orchestration topologies, statefulness and checkpointing, human-in-the-loop gates, runtime guardrails, observability, and controlling cost and latency.
Pattern Catalogue Quiz
Test your understanding of the 18 agent design patterns — goal creation, planning, querying, reflection, cooperation, and safety infrastructure.