Checked August 2026
Agent extension anatomy
Skills teach. Tools act. Plugins extend the runtime.
The vocabulary around agents is often flattened into ‘plugin.’ That hides important trust and architecture differences. A tool is callable. A skill provides procedural context. MCP is a protocol. A plugin installs host-specific runtime code. Memory persists selected state.
Click each layer to inspect its contract, then trace one task through procedure selection, typed tool proposal, transport, policy, execution, and memory.
- Five layers
- Skill · tool · MCP · plugin · memory
- Inspect
- Metadata · schema · transport · handler · state
- Security rule
- Installing instructions is not the same as installing code
Capability stack workbenchComplete path assembled
Procedure
Skill
On-demand instructions and resources tell the agent how to perform a recurring kind of work.
metadata → activate SKILL.md → load references only as needed- Concrete example
- conference-research/SKILL.md activates a source-verification procedure and loads its checklist only when relevant.
- What it cannot do
- A skill cannot call an API or persist data unless the host separately exposes those capabilities.
- Authority
- context read; no inherent execution
Instructions loaded1
Callable capabilities2
Protocol crossings2
Executable packages1
Persistent records0
Primary research
Inspect the mechanism at its source.
Project architecture and terminology were checked against official documentation on August 10, 2026. Research papers are linked to their original publication records.
Open skill formatAgent Skills overviewOriginal source ↗Tools, resources, and promptsMCP server primitivesOriginal source ↗Instruction loading and precedenceOpenClaw skillsOriginal source ↗Runtime extension formatsOpenClaw pluginsOriginal source ↗Procedural memory and approvalHermes Agent skillsOriginal source ↗Tools, hooks, and integrationsHermes Agent pluginsOriginal source ↗