Datasets:

ai-systems
/

ai-system-patterns

pattern stringlengths 11 30	category stringlengths 4 16	description stringlengths 82 114	components listlengths 3 5	use_case stringlengths 19 45	complexity stringclasses 3 values
Model Routing	Orchestration	Routes each request to the model best suited for the task based on capability, latency, cost, modality, or policy.	[ "router", "model registry", "evaluation signals", "fallbacks" ]	Multi-model AI systems	medium
Planner-Executor	Agents	Separates planning from execution so one component creates a task plan while another performs the actions.	[ "planner", "executor", "state", "tools" ]	Long-horizon agent workflows	medium
Supervisor-Worker	Multi-Agent	Uses a supervising agent to delegate specialized subtasks to worker agents and combine their results.	[ "supervisor", "worker agents", "task queue", "result merger" ]	Parallel research and enterprise automation	high
Retrieval-Augmented Generation	Knowledge	Retrieves external knowledge at inference time and injects relevant context into model generation.	[ "retriever", "index", "reranker", "generator" ]	Knowledge-grounded assistants	medium
Tool-Using Agent	Agents	Lets a model select and call external tools such as APIs, browsers, databases, or code execution.	[ "agent model", "tool registry", "schemas", "execution layer" ]	Autonomous task execution	medium
Memory-Augmented Agent	Memory	Adds persistent memory so an agent can reuse relevant information across steps or sessions.	[ "agent", "memory store", "retrieval", "memory policy" ]	Persistent assistants and long-horizon agents	medium
Verifier Loop	Validation	Uses a separate verifier, critic, test, or model to check intermediate or final outputs before acceptance.	[ "generator", "verifier", "retry policy", "acceptance rule" ]	High-reliability reasoning and coding	medium
Human-in-the-Loop	Control	Introduces explicit human review or approval at selected stages of an AI workflow.	[ "agent", "risk policy", "approval interface", "audit trail" ]	High-impact or regulated workflows	medium
World-Model Planning	World Models	Uses a predictive model of the environment to simulate possible outcomes before selecting an action.	[ "world model", "planner", "state representation", "policy" ]	Robotics, simulation, and Physical AI	high
Fallback Routing	Reliability	Switches to an alternate model, provider, tool, or execution path when the preferred path fails.	[ "primary route", "fallback route", "failure detector", "policy" ]	Resilient production AI systems	medium
Multi-Agent Debate	Multi-Agent	Multiple agents propose or critique answers before a final synthesis or decision is produced.	[ "proposer agents", "critic agents", "judge", "shared context" ]	Reasoning and review workflows	high
Agentic RAG	Knowledge	Combines retrieval with agent planning so the system can iteratively search, inspect, and refine evidence.	[ "agent", "retriever", "search tools", "memory", "verifier" ]	Complex research and evidence synthesis	high
Reflection Loop	Reasoning	Lets a model review its own output, identify weaknesses, and produce an improved revision.	[ "generator", "reflection prompt", "revision loop" ]	Reasoning, writing, and coding improvement	low
Self-Consistency	Reasoning	Generates multiple candidate solutions and selects the most consistent result across them.	[ "sampler", "candidate generator", "aggregator" ]	Reasoning and uncertainty reduction	low
Mixture of Agents	Multi-Agent	Combines outputs from multiple specialized agents using a final aggregation or synthesis stage.	[ "specialized agents", "router", "aggregator" ]	Broad multi-domain tasks	high
Event-Driven Agent	Agents	Triggers agent actions from events such as messages, file changes, webhooks, or system state changes.	[ "event bus", "agent", "handlers", "state" ]	Automation and monitoring	medium
State Machine Agent	Agents	Constrains agent behavior to explicit states and transitions for predictable workflow execution.	[ "state machine", "agent", "transition rules", "tools" ]	Reliable structured workflows	medium
Graph-Based Workflow	Orchestration	Represents an AI workflow as nodes and edges so branching, looping, and dependencies are explicit.	[ "workflow graph", "nodes", "edges", "state" ]	Complex agent orchestration	medium
Checkpointed Agent	Memory	Persists task state at checkpoints so long-running work can recover after interruption or failure.	[ "agent", "checkpoint store", "state serializer", "recovery logic" ]	Long-horizon workflows	medium
Hierarchical Memory	Memory	Separates short-term, task-level, and long-term memory to improve context management and retrieval quality.	[ "working memory", "task memory", "long-term memory", "retrieval policy" ]	Persistent agent systems	high
Context Compression	Memory	Compresses older interaction history into summaries or structured state to preserve useful context efficiently.	[ "summarizer", "context store", "retrieval policy" ]	Long-context agent workflows	medium
Permission-Gated Tool Use	Control	Requires policy checks or human approval before an agent can execute selected tools or actions.	[ "tool registry", "permission engine", "identity", "approval layer" ]	Enterprise agents and high-impact automation	high
Sandboxed Execution	Control	Runs model-generated code or actions inside an isolated environment with restricted permissions and resources.	[ "sandbox", "resource limits", "execution engine", "audit" ]	Coding agents and untrusted execution	medium
Observability-First Agent	Observability	Captures traces, tool calls, state transitions, costs, and failures as first-class runtime data.	[ "tracing", "logs", "metrics", "audit store" ]	Production agent operations	medium
Cost-Aware Routing	Orchestration	Chooses models or tools using both capability requirements and explicit cost constraints.	[ "router", "cost model", "capability registry", "budget policy" ]	Cost-efficient AI platforms	medium
Latency-Aware Routing	Orchestration	Selects execution paths based on latency budgets while preserving minimum capability requirements.	[ "router", "latency telemetry", "SLA policy", "fallbacks" ]	Real-time AI applications	medium
Model Cascade	Orchestration	Starts with a cheaper or smaller model and escalates to stronger models only when needed.	[ "small model", "confidence gate", "large model", "router" ]	Efficient inference systems	medium
Ensemble Verification	Validation	Uses several independent models or methods to cross-check a result before accepting it.	[ "candidate output", "verifiers", "aggregator", "acceptance policy" ]	High-confidence AI decisions	high
Synthetic Data Flywheel	Data	Generates synthetic examples, evaluates them, filters them, and feeds high-quality samples back into training.	[ "generator", "quality filter", "deduplication", "training pipeline" ]	Model improvement and domain adaptation	high
Active Learning Loop	Data	Selects uncertain or high-value examples for human or model labeling and uses them to improve the system.	[ "model", "uncertainty scorer", "labeler", "training loop" ]	Efficient supervised data collection	high
Evaluation Harness	Validation	Standardizes datasets, prompts, metrics, and execution settings for repeatable model or agent evaluation.	[ "task suite", "runner", "metrics", "result store" ]	Model and agent benchmarking	medium
Shadow Deployment	Validation	Runs a new AI system alongside production traffic without affecting users, enabling safe comparison.	[ "production system", "candidate system", "traffic mirror", "evaluation" ]	Safe model upgrades	high
Canary Model Rollout	Operations	Deploys a new model to a small portion of traffic before wider release while monitoring quality and failures.	[ "traffic splitter", "candidate model", "monitoring", "rollback" ]	Production model deployment	medium
Human Escalation	Control	Lets the AI system transfer a task to a human when uncertainty, risk, or policy thresholds are exceeded.	[ "risk detector", "escalation policy", "human queue", "handoff context" ]	Customer support and regulated automation	medium
Structured Output Contract	Interoperability	Requires models to produce outputs that conform to a predefined schema for reliable downstream processing.	[ "schema", "validator", "model", "parser" ]	Tool use and software integration	low
Capability Discovery	Interoperability	Allows agents or tools to advertise available capabilities and interface requirements dynamically.	[ "capability registry", "schemas", "discovery protocol", "router" ]	Open agent ecosystems	high
Agent-to-Agent Delegation	Interoperability	Lets one agent delegate a task to another specialized agent while preserving context and expected outputs.	[ "source agent", "target agent", "message schema", "task contract" ]	Distributed multi-agent systems	high
World-State Synchronization	World Models	Keeps the agent's internal representation aligned with changing external environment state.	[ "observation layer", "state model", "update mechanism", "planner" ]	Robotics and dynamic environments	high
Sensor Fusion Pipeline	Physical AI	Combines multiple sensor modalities into a unified representation for perception and downstream decision-making.	[ "sensors", "fusion model", "time synchronization", "state estimator" ]	Robotics, vehicles, and industrial AI	high
Edge-Cloud Split	Physical AI	Splits AI processing between low-latency edge devices and more capable cloud infrastructure.	[ "edge model", "cloud model", "router", "connectivity layer" ]	Robotics, wearables, and mobile AI	high

AI System Patterns

A compact reference dataset of reusable architectural patterns for modern AI systems.

The dataset focuses on practical system-design concepts across AI agents, orchestration, memory, validation, observability, interoperability, world models, Physical AI, data pipelines, and production operations.

Each row contains:

pattern
category
description
components
use_case
complexity

Example

{
  "pattern": "Model Routing",
  "category": "Orchestration",
  "description": "Routes each request to the model best suited for the task based on capability, latency, cost, modality, or policy.",
  "components": ["router", "model registry", "evaluation signals", "fallbacks"],
  "use_case": "Multi-model AI systems",
  "complexity": "medium"
}

Intended Uses

AI system architecture references
taxonomy experiments
lightweight classifiers
retrieval and search demos
documentation examples
agent architecture exploration
educational tools

Limitations

This is a curated reference dataset, not a benchmark and not a comprehensive taxonomy of all AI system architectures. The complexity field is intentionally approximate.

License

Apache-2.0

Maintainer

Published by ai-systems as a practical reference dataset for AI systems and agent infrastructure.

Downloads last month: 10

Number of rows:

Total file size:

17 kB