Orbit is the workspace where humans and agents collaborate — with full visibility, control, and accountability at every step.
The trust model
Every action is logged, named, and surfaced in plain language — no black boxes.
Intent and reasoning are first-class outputs, not optional metadata.
Confidence is declared as a structured signal. Uncertainty is never hidden or smoothed over.
Reversibility is the default. Irreversible actions are explicitly flagged before they happen.
Autonomy is always subject to review. You decide what the agent earns permission to do independently.
Design principles
Every agent action has an attributable explanation path.
Actions are reversible by default. Irreversibility is explicit and rare.
Uncertainty is a first-class signal, never hidden.
New agents start supervised. Trust is built through track record.
Show enough to trust by default. Depth on demand.
Every piece of content traces back to its origin.
External agents have fixed, declared trust properties.
Trust levels
Reads workspace state. Never acts.
Proposes actions. Human approves all.
Executes routine reversible work. Complex actions need approval.
Handles routine work independently. Irreversible actions always need approval.
Features
Every action, summary, and recommendation traces back to its source. Nothing is anonymous. Every piece of output is attributable — to a model, a prompt, a human decision, or a chain of all three.
Agents declare uncertainty as a structured signal. Ambiguity is surfaced, not buried.
Every agent action is either reversible or explicitly flagged as not. No surprises.
Before acting, agents show what they plan to do and why. Humans approve, modify, or reject.
When agents delegate to sub-agents, the full chain is visible and attributable.
"An Orbit capability that increases agent power without increasing human understanding, human control, or human recovery capability is incomplete."
— Orbit Design Principle