Mental model
The six concepts every AgentsKit user should know.
AgentsKit composes around six core concepts. Together they describe everything the framework does. If you understand these six, you can predict how any package behaves.
The six concepts
What each one does
| Concept | One-line definition |
|---|---|
| Adapter | The seam to an LLM provider (OpenAI, Anthropic, etc.) |
| Tool | A function the model can call |
| Skill | A persona the model becomes (system prompt + behavior) |
| Memory | Chat history and vector storage |
| Retriever | Fetches relevant context (RAG, web search, etc.) |
| Runtime | The loop that composes them all |
The two distinctions that matter most
1. Tool vs Skill
A tool is a function. A skill is a persona.
A tool: "give me the weather in Madrid." The model calls it.
A skill: "you are a meticulous code reviewer who insists on tests." The model becomes it.
Confusing them is the recurring failure mode in agent libraries. We make the distinction load-bearing — see ADR 0002 (Tool) and ADR 0005 (Skill).
2. ChatMemory vs VectorMemory
Two different problems. Two different contracts. Two different backends.
ChatMemory is the ordered history of a session. Redis is great at this.
VectorMemory is embeddings indexed for semantic retrieval. pgvector or Qdrant are great at this.
Forcing one interface to do both — as some other libraries do — produces implementations that half-fulfill both sides. We split them.
Where to go next
- Pick a concept page above to dig deeper
- Or jump to the quick start and see them in code