Workspace agents

OpenAI’s new workspace agents in ChatGPT are designed to automate repeatable workflows that involve shared systems, standard handoffs, and consistent outputs. An agent consists of a trigger—such as a schedule or a manual run—, a process that may include specialized skills, and a set of approved tools or integrations it can use to gather information or take actions. These tools can be Slack, a CRM, internal documentation, a ticketing system, or a shared document. The agent’s steps include reviewing inputs, checking for missing information, drafting an output, and handing it off or taking the next action. The output is structured in a clear format so users can assess quality. Workspace agents differ from traditional deterministic API workflows; they are probabilistic, using a model to interpret context and make bounded decisions within instructions and guardrails. Users build agents through an iterative process in the agent builder chat, starting with a plain‑language job description, selecting tools, setting triggers, and adding guardrails such as required approvals or human‑in‑the‑loop checkpoints. The builder translates these inputs into a workflow that can be refined in chat or edited directly. OpenAI provides examples of common agent patterns—such as pulling information from multiple sources, monitoring pipeline changes, or drafting tickets—and suggests starting with low‑risk requests to observe agent behavior. Even well‑built agents benefit from human judgment, and the first version is expected to need refinement. In ChatGPT Enterprise, administrators control access to agent building. The overall goal is to embed AI into day‑to‑day work, moving beyond one‑off tasks to consistent, tool‑connected automation.