Reliable Agent Runtime

Checking public health. Private tasks stay locked until you connect.

Control invariant

The model proposes. The runtime decides.

Tasks, criteria, approvals, tool receipts, checkpoints, and verifier evidence remain durable across runs. A model can propose completion; only the verifier can commit it.

Capability status: connect to inspect configured project workers and credential connectors below. Browser control is not paired to this cloud runtime.

Versioned contract

New task

Select a task to inspect its contract, evidence, actions, approvals, and trace.

Owner-controlled autonomy

Operating controls

Schedules and permissions live on the server. This page does not need to stay open for scheduled work.

Connect to inspect capabilities, schedules, and standing permissions.

Owner controls

Standing permissions

Allow one registered tool for one task or an explicitly scoped project. Grants do not add task permissions or waive existing pending approvals.

Schedules and external events

Enqueue a task once. Approval decisions automatically resume scheduled tasks. An exact event key can wake a task waiting for an external event; it does not prove completion.

Advanced task contract

Create a task with explicit tool-result criteria. No credentials belong in this JSON. A new task is automatically scheduled; unavailable tools or missing permissions remain visible blockers.

Example project and command IDs must match owner-enrolled capabilities.