Primitives

CKP decomposes agent capabilities into 11 primitives. Each primitive is a self-contained, schema-validated building block. Together they describe everything an agent is, knows, and can do.

The 11 primitives

#	Primitive	Kind	Role
1	Identity	`identity`	Who the agent is — personality, context, autonomy level
2	Provider	`provider`	LLM endpoints with fallback chains and token governance
3	Channel	`channel`	Communication surfaces (Telegram, Slack, CLI, Webhook, Voice…)
4	Tool	`tool`	Atomic executable functions (file I/O, web fetch, API call)
5	Skill	`skill`	Composed workflows built from multiple tools with natural-language instructions
6	Memory	`memory`	Persistent storage across sessions plus lifecycle, salience, confidence, and checkpoint semantics
7	WorldModel	`world-model`	Optional predictive planning surface used to simulate consequences before acting
8	Sandbox	`sandbox`	Execution isolation (none, process, WASM, container, VM)
9	Policy	`policy`	Behavioral rules and access control (allow, deny, require-approval, audit-only)
10	Swarm	`swarm`	Multi-agent coordination with 5 topologies
11	Telemetry	`telemetry`	Structured observability events exported from any primitive

How they relate

Primitives form a dependency graph. Some reference others, some compose others, and some observe silently:

Claw Manifest (claw.yaml)
├── Identity (required)          -- who am I?
├── Provider (required)          -- how do I reason?
├── Channel (optional)           -- how do humans reach me?
├── Tool[] (optional)
│   ├── references → Sandbox     -- where does this tool run?
│   └── references → Policy      -- what rules govern this tool?
├── Skill[] (optional)
│   └── composes → Tool[]        -- which tools does this skill use?
├── Memory (optional)            -- what do I remember?
├── WorldModel[] (optional)      -- what futures do I simulate?
│   └── referenced by → Skill[]  -- which workflows use predictive planning?
├── Sandbox (optional)           -- default execution constraints
├── Policy[] (optional)          -- global behavioral rules
├── Swarm (optional)
│   └── references → Identity[]  -- which agents do I collaborate with?
└── Telemetry (optional)         -- what do I export?

Key rules:

Identity and Provider are always required. Every agent must declare who it is and which LLM it reasons with.
Tools bind to Sandbox and Policy. Each tool can reference a specific sandbox for execution isolation and a specific policy for behavioral governance.
Skills compose Tools. A Skill declares which tools it depends on and provides instructions for orchestrating them.
Skills may reference WorldModels. Predictive planning is optional, reusable, and separate from workflow definition.
Swarm references other Identities. Multi-agent coordination involves referencing other agents by their Identity.
Sandbox defaults to most restrictive. If no sandbox is declared, tools run under maximum restrictions.
Policy defaults to deny-all. Unconfigured actions are denied.
Telemetry is purely observational. It never affects agent behavior.

Primitive addressing

Every primitive instance is addressable via a claw:// URI:

claw://local/identity/research-assistant
claw://local/tool/[email protected]
claw://registry/community-skills/[email protected]

This enables tools, skills, and swarm peers to reference each other unambiguously. See URI Scheme for the full grammar.

Mapping to conformance levels

The primitives map directly to the three conformance levels:

Level	Required Primitives	Use Case
L1 Core	Identity, Provider	Simple chatbot — receive input, reason, respond
L2 Standard	+ Channel, Tool, Sandbox, Policy	Interactive agent with tool execution and security
L3 Full	+ Memory, Skill, Swarm	Autonomous agent with persistence, composition, and collaboration

WorldModel and Telemetry are optional at all levels. For the full conformance breakdown, see Conformance Levels.

Individual primitive specs

Each primitive has its own detailed specification page:

Identity — Agent personality, system prompt, autonomy
Provider — LLM endpoints, fallbacks, token limits
Channel — 16 channel types, access control
Tool — Function schemas, execution model
Skill — Workflow composition, tool dependencies
Memory — 5 store types, lifecycle, salience, confidence
WorldModel — predictive planning, backend selection, bounded rollouts
Sandbox — 5 isolation levels, resource limits
Policy — Rule engine, prompt injection detection
Swarm — 5 topologies, delegation, trust
Telemetry — Exporters, sampling, structured events