How agents work

CORA is the front door to the work in your workspace. You talk to her; she answers what she can herself and hands larger pieces of work to specialist agents.

CORA and her specialists

CORA handles questions, memory, settings and everyday requests directly. For domain work such as research, marketing, sales or finance, she can delegate to a specialist. Specialists work behind CORA rather than as separate things you manage, and a specialist may be named in a task's activity. A specialist can pass part of a job on, but delegation stops at three levels, for example CORA, then Marketing, then Researcher.

Tasks

Every durable piece of work becomes a task. A task has a status, records who asked for it and who approved it, and keeps each step it took. Tasks start from a chat, from a workflow, or when one agent delegates to another.

Agents do not change things directly. When an agent wants to send an email, update another system or save something, it proposes a task for that action, and the action follows its own trust level. See Approvals and trust levels.

Live progress

Delegated work appears in the chat as a live card. The agent first sets out a plan, then each step shows its status, such as Not started, Under way, Done, Needs your approval, Skipped or Failed. The progress is stored as it happens, so the card is still current if you close the browser and come back. When the task finishes, the card shows the result with a link to open it.

Finished work is kept as a work item: the report, draft, brief or summary. Use Pin to knowledge on a finished item to make it part of your company context.

Plans before work

By default, delegated work starts without a separate plan approval. This setting is called Agent's choice. The specialists' own tools only read, and every write still waits at its own approval. A workspace that wants to review every plan can choose Always ask under Workspace settings -> Agent permissions -> Before work begins. When a plan is waiting, you can approve it, decline it or ask for a revision in the chat.

Bounded runs

Each agent run is short and has limits on time and size, so a run that goes wrong stops instead of running on. An agent that delegates does not sit waiting for the other agent; its run ends, and the delegated work continues as its own task. Every plan, step, source and approval is recorded.