An agent is a persistent, reusable AI persona with its own brain and tools. Configure it once — identity, instructions, model, knowledge, and capabilities — then use it across chat, workflows, and the API, or have it coordinate a team of sub-agents.
Unlike a one-off chat prompt, an agent is a saved configuration you can invoke anywhere. Each agent carries its own model, a system prompt, conversation starters, attached knowledge folders, a set of tools and integration actions it may call, optional sub-agents, and an optional structured-output schema. You build it once, share it, version it, and reuse it — in chat, as a node in a workflow, behind the API, or as a fillable form.
Agents run on the autonomous ReAct loop: the agent reasons about your request, decides which tool or knowledge source to call, observes the result, and repeats until it can answer. That reason → act → observe cycle is what makes an agent more than a single prompt.
Your Office AI gives you three building blocks for AI behaviour. They are easy to confuse, so here is the clear line between them — pick the right one for the job:
| Building block | What it carries | How it behaves | Use it when… |
|---|---|---|---|
| Agent | Its own model, system prompt, knowledge folders, tools, and sub-agents | Autonomous persona — reasons and decides which tools to call | You want a capable, reusable assistant that can act. |
| Skill | An instruction snippet only — no model, no knowledge, no tools | Steers an existing model's behaviour when toggled on | You just want to shape tone or rules. Skills. |
| Workflow | A deterministic graph of nodes with explicit wiring | Runs the same path every time — branching, looping, predictable | You want a repeatable, multi-step process. Workflows. |
The three compose: a workflow can contain an agent node, and an agent can attach skills and delegate to sub-agents. Skills only steer behaviour, so when you need a model, knowledge, or tools, reach for an agent.
Agents execute on the LangGraph ReAct runtime — a Python sidecar fronted by a Dart AI server that handles authentication, budgets, and event streaming. The same engine powers workflows, so agents and workflows share one execution backbone. Human-in-the-loop approval (the write-action gate) is implemented natively with LangGraph's interrupt mechanism.
This section walks through everything you can do with an agent, from building one to delegating across a team:
A guided five-step builder — identity, system prompt, conversation starters, model and capabilities, then sub-agents and form fields.
Give an agent integration actions, web search, knowledge retrieval, and image generation — and control the write-action confirmation gate.
Attach knowledge folders for retrieval, give the agent memory, and define a structured-output schema so it returns predictable JSON.
Compose a team of specialist agents behind one entry point with the supervisor delegation pattern.
Mention an agent in chat, drop it into a workflow, call it over the API, or run it as a form — plus templates, the Try panel, and versions.
Every agent picks its own model from the providers your administrator has configured. Your Office AI supports six providers in the same organization — hosted models from OpenAI, Anthropic, Google, and Groq, plus private, self-hosted models through Ollama and Ollama Cloud. The model picker is data-driven from your organization's configuration, with current models such as claude-sonnet-4-5, gpt-4o-mini, gemini-2.0-flash, Ollama llama3.2, and Ollama Cloud gpt-oss:120b.
Agents use the same four sharing scopes as the rest of Your Office AI — Private, Specific people & Agents, Workspaces, and Organisation — each with view or edit. Members & Roles.