Skills, tools and execution

The sidebar's Concepts glossary uses these same terms. They describe separate choices, not alternative names for one capability.

Term	Meaning
Agent	Identity and configuration; not proof that compute is running.
Skill	Instructions, domain knowledge or a workflow the agent follows.
SkillPack	Packaged skills, optionally with sidecar tools, dependencies and RBAC.
Tool	An operation the agent can invoke. Instructions alone do not grant access.
Harness	The model/tool-use loop. AgentRuntime describes an approved runtime; AgentHarness is not a separate resource here.
Backend	Where execution occurs: the Kubernetes pod path or Celln.
Lifecycle	One-shot or enduring execution; independent of harness choice.
AgentRun	The execution record, including lifecycle, backend, status and results.
AgentRunTurn	A follow-up message and result within an enduring native run.

A Kubernetes diagnostic skill explains how to investigate. Its tools perform the actual cluster reads. RBAC or native grants determine which reads are allowed. A skill cannot bypass those permissions.

Compatibility today

Existing SkillPacks such as k8s-ops, GitHub GitOps and SRE observability remain available on their supported paths. They are not automatically converted to borrowed Celln tools. The native starter rejects run-level SkillPacks and does not provide arbitrary MCP, shell, Python or Kubernetes administration. A skill may itself assume pod credentials or a sidecar; copying its text is not parity.

The native starter offers explicitly selected, reviewed revisions of:

• workspace-read: bounded logical files belonging to the live run.
• workspace-write: bounded, revision-checked updates to those files.
• https-fetch: bounded host-brokered HTTPS to approved destinations (the first starter profile permits example.com).

These are not arbitrary host filesystem access or unrestricted shell commands. The operator, runtime and agent grants must all permit access. A permission preview explains grants; it neither issues authority nor proves execution readiness. Tool selection is fixed for a live native parent.

Persistence is backend-specific

Native Celln uses AgentRun.spec.executionLifecycle: enduring: one leased parent cell and a disposable child (sub-cell) per turn. A child is a separate cell, not an AI sub-agent or nested VM. Context is currently approximately 2 KiB. Run-owned files and live context disappear with the parent; checkpoint/resume after parent or host loss is not supported. Controller/API restarts do not themselves destroy the host parent. Saved history is not a restorable VM.

Existing OCI HarnessSession chat is different. Supported adapters may retain state on a PVC and implement stop/resume; that does not transfer to Celln. One-shot Celln work remains disposable and can invoke a deterministic tool with no model or harness at all.

Choosing through UI or YAML

Choose a runtime compatible with the backend. Native YAML selects the runtime and tool revisions in cellnSelection, and lifecycle/ceilings in executionLifecycle and enduring. The operator template must match the persona, model, tools and limits. See native installation.

The run form blocks native selection for an Agent with SkillPacks. The API and admission webhook also reject SkillPacks and Agent MCP connections explicitly; use the dedicated installed native starter Agent. Incompatible selections must not silently drop skills or switch backends. Full installed acceptance is still being qualified; this is not a capability-parity claim.