← Glossary · Concept
Async agents — work that does not live in chat
An agent that takes a task, runs outside the conversation, and returns a finished artifact. Powerful for long workflows, dangerous without state, limits and review.
What it is
An async agent is agentic work detached from a live chat. You do not wait for the model to finish typing. You give it a goal, it gets an environment, runs for minutes or hours, stores state and returns an artifact: a pull request, a report, a draft article, a data export or a decision list.
That is a real shift. Chat is a conversation. An async agent is a work queue.
When it makes sense
It fits tasks that need exploration, tools and time: reading a repository, connecting source material, drafting a change, processing a batch of data or watching new signals and sending them for approval. A human should not sit beside a token metronome pressing “continue”.
What needs guardrails
An async agent needs hard boundaries: which tools it may call, how many steps it may take, where it stores state, what is public and what is only a draft. Without that, convenient automation becomes an expensive black box with production access. Hellishly practical, hellishly risky.
Difference from ordinary automation
A cron job or workflow engine follows predefined steps. An async agent chooses part of the path based on what it finds. That is useful when you do not know exactly what it will discover. It is also why you need logs, previews and human approval before exposing the result or deploying it.
What to remember
An async agent is not “a chatbot that thinks longer”. It is a delegated unit of work. It works best with a narrow mandate, an isolated environment, persistent state and a clear definition of done.