Human-in-the-loop

Full automation is exactly what working scientists distrust — a system that runs off and does things is a liability when the stakes are your name on a paper. CoSci is built as AI-in-the-loop: it augments you, it doesn’t replace you.

Three task classes

Every unit of work CoSci proposes falls into one of three classes, so you always know what it will and won’t do on its own:

• needs_input — CoSci is waiting on a decision or information from you.
• needs_approval — CoSci has prepared an action but will not run it until you approve.
• ai_can_execute — safe, reversible work CoSci can carry out on its own.

The Home screen surfaces everything in the first two classes as what needs you.

Sensitive actions are enqueued, not executed

Actions that change important state or reach outside your machine are enqueued for your approval rather than executed automatically. Nothing consequential happens silently — you review it first, and every decision is recorded.

Everything is visible

The Agents / Runs screen logs every model and tool call, with the provider it used, cache state, and online/offline status. Combined with the approval gates, this means there is no off-screen behavior to trust blindly — you can audit exactly what happened and why.

This pairs with local-first privacy: your data stays on your machine and nothing acts on it without you.