Conformance Levels

Not every agent needs every feature. CKP defines three conformance levels so implementations can start small and grow. A 4 MB embedded chatbot and a 52-module enterprise swarm are both valid CKP agents — they just conform at different levels.

---

Level 1: Core (Minimum Viable Claw)

The smallest possible CKP agent. Receives input, reasons with an LLM, and responds.

Required primitives: Identity, Provider

Capabilities:

• Parse and validate a claw.yaml manifest
• Initialize with claw.initialize and report status with claw.status
• Reason with at least one LLM Provider
• Respond to input via an implicit CLI channel (stdio)
• Shut down gracefully with claw.shutdown

Cannot:

• Execute tools
• Persist memory across sessions
• Coordinate with other agents
• Connect to external messaging channels

Use cases: Embedded devices, simple chatbots, prototyping, CI pipelines.

Test vectors: 13 (see test vectors)

---

Level 2: Standard (Interactive Agent)

An agent that connects to the outside world, executes tools, and enforces security.

Required primitives: Identity, Provider, Channel, Tool, Sandbox, Policy

Adds over Level 1:

• Messaging channels (Telegram, Slack, Discord, Webhook, etc.)
• Tool execution within sandbox constraints
• Policy rules (allow, deny, require-approval, audit-only)
• Approval gates for sensitive operations
• Channel access control (allowlists, authentication)

Use cases: Personal assistants, team bots, automated workflows, customer support agents.

Test vectors: 10 additional (23 cumulative)

---

Level 3: Full (Autonomous Swarm Agent)

A fully autonomous agent with persistence, composed workflows, and multi-agent collaboration.

Required primitives: All 9 core primitives (WorldModel and Telemetry optional at all levels)

Adds over Level 2:

• Persistent memory (conversation, semantic, key-value, workspace)
• Skill composition (multi-step workflows from tools)
• Swarm coordination (5 topologies: star, mesh, pipeline, hierarchical, broadcast)
• Memory compaction and semantic search
• Cross-agent delegation and task reporting
• Optional predictive planning through reusable WorldModel declarations

Use cases: Enterprise deployments, research teams, complex multi-step workflows, autonomous operations.

Test vectors: 8 additional (31 cumulative)

Those 8 L3 vectors are a mix of live runtime behavior and manifest validation:

• live runtime coverage: swarm delegate/report, broadcast, discover, memory store/query, and compact
• manifest-validation coverage: invalid Channel.access_control combinations for allowlist and role-based modes

That split is intentional. CKP conformance includes both wire-protocol behavior and schema-level correctness where the protocol depends on manifest structure.

WorldModel does not add wire conformance requirements in CKP 0.3.0. It is validated first at the manifest/schema layer and described operationally in the Runtime Profile.

---

Method support matrix

Method	Level 1	Level 2	Level 3
claw.initialize	MUST	MUST	MUST
claw.status	MUST	MUST	MUST
claw.shutdown	MUST	MUST	MUST
claw.heartbeat	SHOULD	SHOULD	MUST
claw.initialized	SHOULD	SHOULD	SHOULD
claw.tool.call	—	MUST	MUST
claw.tool.approve / deny	—	MUST	MUST
claw.swarm.delegate	—	—	MUST
claw.swarm.report	—	—	MUST
claw.swarm.broadcast	—	—	MUST
claw.swarm.discover	—	—	MUST
claw.memory.query	—	—	MUST
claw.memory.store	—	—	MUST
claw.memory.compact	—	—	MUST

---

Declaring conformance

Conformance level is not declared in the manifest — it is inferred from the primitives present. A runtime validates the manifest against the requirements for each level and reports the highest level satisfied.

# This manifest is Level 1 (only Identity + Provider)
name: simple-bot
version: "1.0.0"
identity:
  system_prompt: "You are a helpful assistant."
  autonomy: supervised
providers:
  - name: claude
    kind: anthropic
    model: claude-sonnet-4-20250514

# This manifest is Level 2 (adds Channel, Tool, Sandbox, Policy)
name: team-assistant
version: "1.0.0"
identity:
  system_prompt: "You help the engineering team."
  autonomy: supervised
providers:
  - name: claude
    kind: anthropic
    model: claude-sonnet-4-20250514
channels:
  - name: slack
    kind: slack
    access_control:
      mode: allowlist
      identifiers: ["U12345"]
tools:
  - name: web-fetch
    description: "Fetch a URL"
    input_schema: { ... }
sandbox:
  level: process
policies:
  - name: standard
    rules:
      - pattern: "tool:*"
        action: require-approval

---

Choosing a level

Question	L1	L2	L3
Does it just chat?	Yes	—	—
Does it call APIs or tools?	—	Yes	—
Does it need to remember things?	—	—	Yes
Does it work with other agents?	—	—	Yes
Is it embedded / resource-constrained?	Yes	—	—
Does it need audit trails?	—	Yes	Yes

Start at Level 1 and promote when you need the capabilities of a higher level. The dependency graph ensures you add primitives in the right order.