SCREENS / GROUP
Assistants and agents held to the same rules as the numbers: every answer shows its sources, every action waits for a person, every run is readable step by step.
01 / ai-001
A conversation built to be checked: each answer leads with its sentence, then a chart in the system's own rules, then — folded — every call it made and the rows it read, with a projection labelled as one.
src/app/(screens)/_components/screens/ai/ask-stella.tsxdesktop default02 / ai-002
What the agents did unattended, as a queue in four visibly different states, and one run as its steps: plan, each tool call with its tokens, the live step with its clock, and a Stop that says what it keeps.
src/app/(screens)/_components/screens/ai/agent-runs.tsxdesktop default03 / ai-003
Nothing an agent proposes changes the business until a person says so: the change as a before and after, the figure that convinced it, the risk in a word, when the offer lapses — and Undo after deciding.
src/app/(screens)/_components/screens/ai/approvals.tsxdesktop default04 / ai-004
An agent defined in the order that limits risk: purpose, a read / propose / act level per tool with the consequence of act beside it, the data it may read, a spend cap — and a bench to test the draft.
src/app/(screens)/_components/screens/ai/agent-setup.tsxdesktop default05 / ai-005
What the agents cost against a budget line and whether they are still right: a fixed evaluation suite per agent, re-run on every change, with a regression named and the change that caused it.
src/app/(screens)/_components/screens/ai/usage-evals.tsxdesktop default