The vocabulary of agent construction
Every agent is assembled from two kinds of primitive: knowledge shapes (how the agent represents what it knows, the world model) and functional nodes (what a part of the agent does, plan, route, tool, memory). Ontologies decide how it thinks; node types are the mechanics.
| Ontology | Mechanism | Example | src | |
|---|---|---|---|---|
| Schema-as-ontology (JSON Schema / OpenAPI) Schema | Tools and actions are defined as typed JSON Schema / OpenAPI: the agent knows each tool's inputs, outputs and constraints from the schema itself. | AWS Bedrock Action Groups | ↗ | |
| MCP capability primitives Protocol primitives | A fixed vocabulary of capability types, Tools, Resources, Prompts, Sampling, Elicitation, Roots, each with JSON-RPC methods (tools/call, resources/read, prompts/get…). The shared ontology of what a server can offer an agent. | Model Context Protocol | ↗ | |
| Temporal knowledge graph Knowledge graph | Facts are nodes; relationships are edges carrying valid_from / invalid_at metadata: a time-aware model of the world the agent can query for "what is true now vs then". | Zep / Graphiti | ↗ | |
| N-ary hypergraph (AMR) Knowledge graph | Abstract Meaning Representation with n-ary hyperedges: captures relations richer than subject-predicate-object triples. | Hypabase | ↗ | |
| Typed runtime state Typed state | The agent's working state is a typed schema (TypedDict / Pydantic) the graph reads and writes at each step: a structured, inspectable model of the run. | LangGraph | ↗ | |
| Multi-agent conversation schema Typed state | Roles, turns and termination conditions of a multi-agent chat are defined by a schema the orchestrator enforces. | AutoGen | ↗ | |
| Implicit toolbelt (ReAct) Implicit | No formal ontology, the model reasons over natural-language tool descriptions injected into the prompt. Flexible, but the "schema" lives only in the LLM's interpretation. | ReAct | ↗ | |
| Domain reasoning engine Domain reasoning | The Atlas Reasoning Engine + Data Cloud + Apex @InvocableMethod expose enterprise objects and actions as the agent's ontology: it reasons in the business's own data model. | Salesforce Agentforce | ↗ | |
| Memory-type ontology Memory typing | Semantic / episodic / procedural typing structures what the agent retains and how it is retrieved: the ontology of the agent's own experience. (See Software → Memory.) | LangMem / Mem0 | ↗ | |
| Governed enterprise ontology (Palantir Ontology) Digital twin | A governed digital twin of enterprise data, business logic and compliance rules. Two layers: the SEMANTIC layer (Object Types, Link Types, Properties, Structs, Shared Properties, Value Types, Interfaces) models real-world entities and their relationships; the KINETIC layer (Action Types, Functions, Function-Backed Actions) captures how state can change under governance. A DYNAMIC/security layer (Object Security Policies, Object Permissioning) enforces the same compliance guardrails on agents that human operators face. Agents traverse object links to gather context, execute predefined Action Types instead of touching data directly, and compose LLM reasoning with deterministic simulation/optimisation tools wired in as Functions. Distinct from schema-as-ontology (JSON Schema / OpenAPI): those describe tool signatures; this models the entire enterprise. | Palantir Foundry / AIP | ↗ |
| Node | Function | Example agents | |
|---|---|---|---|
| Intent / Classification | routes a request to the right skill or sub-agent | Intercom Fin, Agentforce | |
| Planner / Decomposition | breaks a goal into ordered steps | Devin, Claude Code | |
| Tool-calling | invokes external tools / APIs | most agents | |
| Context Retrieval (RAG) | fetches relevant context from a store | Notion, Glean | |
| Web Search | queries the live web | ChatGPT, Perplexity | |
| Code Execution | runs generated code in a sandbox | Devin, ChatGPT | |
| Browser / Computer-use | drives a real browser or desktop | OpenAI Operator, Claude | |
| Memory Read/Write | persists and recalls state across turns | Lindy, Mem0 users | |
| Router / Orchestrator | routes between sub-agents / models | Agentforce, Cursor | |
| Validator / Guardrail | checks outputs against policy/safety | enterprise agents | |
| Human-in-the-loop | pauses for human approval | high-stakes enterprise agents | |
| Formatter / Output | shapes the final structured response | all agents |
Ontologies decide how it thinks; node types are the mechanics
Building blocks are the primitives an agent is assembled from, the units below a topology and above a library: a planner, a router, a critic, a memory read, a tool call, a gate. Naming them is what makes two systems comparable.
Show more
Most production designs use a handful, and the interesting variation is in how they are wired rather than which are chosen. Once the vocabulary is fixed, an architecture stops being a diagram and becomes something you can reason about, test in pieces and change one part at a time.
What are the building blocks of an AI agent?
A handful of parts you assemble: a planner that works out the steps, a router that picks which way to go, tools it can call, memory it reads and writes, and validators or gates that check output before it is accepted. Most agents use only a few of these.
What does a planner do in an AI agent?
It turns a goal into an ordered set of steps before any of them run. Better planners re-plan partway through when a step returns something unexpected. Without one, an agent only reacts turn by turn and tends to lose the thread on long tasks.
What is a router in an agent system?
A router looks at an incoming request and sends it to the right place: the right sub-agent, the right tool, or the right model. It is how a system uses a small cheap model for easy requests and saves the expensive one for the hard ones.
What is a guardrail or validator in an agent?
A check that runs on output before it is accepted, usually against a schema or a rule. If the check fails the agent retries or escalates instead of passing bad data downstream. A gate is the stricter version: it stops and waits for a person.
What is the difference between a tool and an agent?
A tool does one defined thing when called: query a database, send an email, run some code. An agent decides which tools to call and in what order. An agent can also be exposed to another agent as a tool, which is how sub-agents are usually wired in.
What is an agent ontology?
The shape of what the agent knows: a fixed schema, a knowledge graph, typed state, or nothing explicit at all. It decides what the agent can reason about precisely and what it can only infer from text. Most agents leave it implicit, which is fine until two facts contradict each other.