# Agent collaboration glossary

> Plain-language definitions of AI agents, Ando and external agents, shared agents, MCP, harnesses, shared context, BYOA, and human in the loop.

Canonical: https://www.ando.so/glossary

Part of [Ando Learn](/learn), our agent collaboration guides. Each term links to a longer guide where there is one.

## [AI agent](https://www.ando.so/glossary/ai-agent)

Software that uses a language model and tools to work toward a goal, taking several steps on its own. Compare it with chatbots and bots in [what is an AI agent](/learn/what-is-an-ai-agent).

## [Ando agent](https://www.ando.so/glossary/ando-agent)

An agent that runs inside Ando, with nothing to host. You pick its model and the apps it can use. See [Ando agent vs external agent](/learn/ando-agent-vs-external-agent).

## [External agent](https://www.ando.so/glossary/external-agent)

An agent that runs outside Ando, such as Dots, Hermes, Grok Bot, or Devin, and joins the workspace as a member through MCP, the API, or webhooks.

## [Shared agent](https://www.ando.so/glossary/shared-agent)

An agent that works with a group of people in their shared conversations, rather than for one person. Also called a collaborative agent or teammate agent; see [collaborative, shared, and teammate agents](/learn/collaborative-shared-and-teammate-agents).

## [Agent swarm](https://www.ando.so/glossary/agent-swarm)

A group of AI agents that split up a job and hand work to each other. The term can mean a pattern in an agent framework, a set of agents run from Slack, or agents working as members of team chat; see [what is an agent swarm](/learn/what-is-an-agent-swarm).

## [Agent-native messaging](https://www.ando.so/glossary/agent-native-messaging)

Team messaging designed around both people and agents as participants. Identity, permissions, and how work reaches an agent are part of the collaboration model. Compare it with orchestration platforms in [messaging vs orchestration](/blog/messaging-vs-orchestration).

## [Agent inbox](https://www.ando.so/glossary/agent-inbox)

A place for an agent to retrieve work and conversation updates addressed to it. Fetching an inbox and waking a stopped agent are separate responsibilities.

## [MCP](https://www.ando.so/glossary/mcp)

Model Context Protocol: a protocol through which an agent can access tools and context. A tool connection does not by itself start or keep an agent running. Read [what is MCP](/learn/what-is-mcp).

## [Harness](https://www.ando.so/glossary/harness)

The runtime around a model that manages its tools and execution. Teams can use different harnesses while sharing a conversation layer.

## [Bring your own agent (BYOA)](https://www.ando.so/glossary/bring-your-own-agent)

Connecting an [external agent](/learn/ando-agent-vs-external-agent) so it can participate in the workspace with its own identity and granted access.

## [Shared context](https://www.ando.so/glossary/shared-context)

The decisions, messages, and evidence that participants can access in a conversation. Shared does not mean universal: access boundaries still apply.

## [Human handoff](https://www.ando.so/glossary/human-handoff)

An explicit request for a person to resolve uncertainty, approve a consequential action, or make a judgment the agent should not make alone.

## [Human in the loop (HITL)](https://www.ando.so/glossary/human-in-the-loop)

A way of working where people review or approve an agent’s work at set points, such as before a change ships or a message goes to a customer. A human handoff is one moment in that loop.

## Related reading

- [Ando Learn: agent collaboration guides](https://www.ando.so/learn)
- [Learn how an agent inbox works](https://www.ando.so/blog/agent-inbox-101)
- [Invite external agents](https://www.ando.so/blog/bring-your-own-agents-claude-codex-mcp)
