Skip to content

Core Concepts

A mesh is the set of all agents and tools registered with a single AgentField control plane. Agents find and call each other dynamically — no hardcoded references.

An agent is an AI-powered service that:

  • Registers on the control plane
  • Has a system prompt that defines its behavior
  • Has skills (prompt instructions) that reference connectors (tools)
  • Discovers other agents and tools at runtime via tools="discover"

A skill is a prompt instruction — it tells the agent’s LLM how to behave and what tools to use. Skills are pure markdown, no code.

skills/fast-analytics-query/SKILL.md

Skills reference connectors by name. When an agent is scaffolded, skill instructions are appended to the system prompt.

A connector is a deterministic function — it calls an external API, processes data, or reads/writes memory. No LLM involved.

tools/fast-query/
TOOL.md # Metadata
impl.py # Async Python function
config.yaml # Environment-aware config + secrets

Teams group related agents into a sub-mesh with namespaced memory. Agents in a team share state with each other but not with the parent mesh.

Agent = System Prompt + Skills + Model
└── Skills reference Connectors (resolved transitively)
└── Connectors are discovered at runtime via control plane