How Agents Work
What an agent is
A server-side agent is a workspace-scoped AI participant defined by its configuration: a name and handle (its @mention, e.g. @plan-bot), an avatar, an LLM provider and model, an optional system prompt, and a few limits. Agents belong to the workspace, so the same agents are available everywhere in it.
Where agents take part
An agent's handle appears in the mention picker on every discussion surface:
- a task's Discussion tab,
- a brainstorm,
- a channel,
- the AI Assistant panel.
Type @, pick the agent, and write your request.
What a run does
Mentioning an agent starts a run:
- The agent receives the recent thread context.
- It makes one or more requests to its LLM provider.
- Between requests it may call tools — for example list or update tasks, search the knowledge base, or read a project — exactly the operations a person can do, within the triggering user's permissions.
- It posts its final answer as a normal message in the thread.
Because agents use the same tools and permission checks as the app, they are genuine collaborators: they can pick up a task, read its context, make changes, write results into the knowledge base, and discuss — all recorded in the activity history under the person who triggered them.
Three complementary kinds of agent
| Server-side agent | Harness agent | Local agent | |
|---|---|---|---|
| Lives | In the workspace (Settings → Agents) | In the workspace, working on a runner | On your machine (Claude Code or another MCP client) |
| You reach it by | @mention in a thread |
@mention in a thread |
Your editor / terminal, over MCP |
| Has a computer | With OpenAI and Anthropic models — the provider's own sandbox for running code, with its limits; other providers — no | Yes — its own sandbox with a shell, files and internet | Yours |
| Best for | Quick, in-context help right where the conversation is — summarize a thread, draft a subtask breakdown, create tasks from a discussion, or answer a question from the knowledge base | Work that needs a computer — analysing spreadsheets and documents, real calculations and charts, collecting data from the web, producing files, multi-step jobs that build on earlier results | Multi-step, hands-on work from your editor — plan and implement a task end to end, or make bulk changes across the workspace |
The next pages cover configuring an agent, using it, and what makes harness agents different. Local agents — including the /task-plan and /task-implement skills — are in Work from your own AI (MCP & Skills).