Agentic Primitives
The Building Blocks of AI Agent Systems
👤
Actors
Who participates?
Users
Reference
Human participants who initiate tasks, provide context, and consume outputs.
Agents
Reference
AI entities that perceive, reason, and act to accomplish goals.
🔧
Tools
What can agents access?
Knowledge Tools
LLM + Platform
Read-only interfaces. LLM uses descriptions; Platform handles bindings.
Vector Search
Database Query
Web Search
Action Tools
LLM + Platform
Interfaces that modify external state. LLM decides; Platform executes.
API Mutation
Code Execution
Email Sender
📋
Instructions
What guides behavior?
Agent Instructions
LLM
Core personality, expertise, tone injected into prompts.
Workflow Instructions
LLM + Platform
Step-by-step procedures. LLM follows; Platform may encode.
System Instructions
LLM
Organization-wide policies: constraints and intent.
🎯
Coordination
How is work coordinated?
Workflow Orchestration
Centralized
Platform
Deterministic sequences in workflow engine.
Agentic Orchestration
Centralized
LLM + Platform
LLM-driven coordination. Platform provides runtime.
Choreography
Decentralized
Platform
Event-driven via event bus and pub/sub.
🔗
Connections
How do components link?
Point-to-Point
Platform
Direct connections. URLs, auth, protocols in infrastructure.
Dynamic
Platform
Dynamic lookup from registries and catalogs.
Queued
Platform
Asynchronous message queues for buffered, reliable delivery.
💬
Interactions
What flows between participants?
Delegation (Command)
LLM + Platform
Instructions requesting action. Platform routes; LLM processes.
Retrieval (Query)
LLM + Platform
Requests for information. Platform routes; LLM responds.
Notification (Event)
Platform
State change announcements via webhooks, event bus.
Conversations
LLM + Platform
Sustained exchanges. Platform manages threads; LLM uses history.
LLM
Consumed by the language model in prompts/context
Platform
Configured in orchestration platform/infrastructure
LLM + Platform
Components for both
Reference
Documentation/governance