GenAI Field GuideAgents and beyond
Tools, connectors (MCP), and skills
A tool is a function the harness exposes to the model with a name, parameters, and a description. The model emits a structured request; the harness executes it; the result returns as text in the context. The model never touches your files directly.
The three words you will hear
| Term | What it is | What it means for you |
|---|---|---|
| Tool | One function with a schema: search, run code, read a file, check a calendar | If the assistant "can't do" something, it probably lacks the tool, not the ability |
| MCP (Model Context Protocol) | The now-universal standard for plugging tools into any model. Anthropic proposed it; OpenAI, Google, and xAI adopted it. "Connectors" in consumer products are MCP servers. | Every connector you add becomes more text in the prompt and more ways the agent can act |
| Skill | A folder of instructions plus scripts the harness loads on demand: "how we write a Navy memo," "how to build a slide deck" | The cheapest way to make output match your office's format, and it lives in the preamble layer |
Two facts that explain most surprises
- Adding a tool changes behavior with no change to the model. Same model, new connector, different assistant. When one product "can" and another "can't," look here first.
- Tool results are truncated. Harnesses cut large outputs (the head and tail of a log, the first N kilobytes of a file) before appending them, so the model may not have seen the whole thing.