Authority and autonomy
Autonomy is not a setting on a model; it is authority granted to an agent for a workflow, written as an envelope Alice enforces on every decision. Three levels describe how much the agent may do before a human is in the loop. Adapted from the three phases in Microsoft's 2025 Work Trend Index (human with assistant · human-agent teams · human-led, agent-operated); the adaptation is Nexus's own and implies no partnership.
Three levels
Human with assistant
The agent drafts, summarizes, and suggests. A human performs every action that changes a record or leaves the boundary.
Human role: Performs the action; reviews every output.
Human-agent team
The agent performs bounded internal actions inside its envelope. Anything with external effect, elevated risk, or irreversibility is held for a human.
Human role: Directs the work; approves exceptions.
Human-led, agent-operated
The agent runs a workflow end to end inside a tightly specified envelope. Humans set direction, receive receipts, and check in on exceptions.
Human role: Sets direction; audits receipts; intervenes on exceptions.
An organization is in all three levels at once across different workflows. Level is granted per agent and workflow, reviewed on a schedule, and lowered automatically by readiness signals (see the GOV-04 cap on the command center).
What each level may do
| Action | L1 · Assist | L2 · Team | L3 · Operate |
|---|---|---|---|
| Draft content | Agent drafts · human acts | Inside envelope | Inside envelope |
| Summarize | Agent drafts · human acts | Inside envelope | Inside envelope |
| Classify | Agent drafts · human acts | Inside envelope | Inside envelope |
| Analyze | Agent drafts · human acts | Inside envelope | Inside envelope |
| Write to a system of record | Human performs | Inside envelope | Inside envelope |
| Execute a tool with side effects | Human performs | Inside envelope | Inside envelope |
| Send outside the boundary | Human performs | Held for approval | Held for approval |
| Publish externally | Human performs | Held for approval | Held for approval |
Always, at every level: an action outside the envelope is blocked; an irreversible action needs approval; risk at or above the envelope threshold is held; PHI never leaves the boundary; a receipt is issued.
The autonomy envelope
Every agent carries one. Alice reads it on every decision; the agent cannot change it.
Which Nexus classes (Alice L1–L4) the agent may see.
Draft, summarize, classify, analyze, write a record, execute a tool, send outside, publish.
Named tools only; anything else is refused (the kernel's no_tools restriction is an empty list).
A subset of the governed registry; model eligibility is still evaluated per request.
Actions that always need a named approver, a risk threshold, and irreversibility.
Receipt always; reviewer recorded; retention period.
Envelopes in the registry
| Agent | Level | Data | Actions | Approval required for | Risk threshold | Approver |
|---|---|---|---|---|---|---|
| Discharge Summary Assistant | L1 · Human with assistant | PUBLIC, INTERNAL, PHI | Draft content · Summarize | Draft content · Summarize · irreversible | ≥ MEDIUM | Attending physician |
| Scheduling Optimizer | L2 · Human-agent team | PUBLIC, INTERNAL | Analyze · Draft content · Write to a system of record · Send outside the boundary | Send outside the boundary · irreversible | ≥ HIGH | Director of Operations |
| Nexus Intelligence Analyst | L3 · Human-led, agent-operated | PUBLIC | Analyze · Summarize · Draft content · Execute a tool with side effects · Publish externally | Publish externally · Send outside the boundary · irreversible | ≥ HIGH | AI Governance Lead |
| Staff Drafting Assistant | L1 · Human with assistant | PUBLIC, INTERNAL | Draft content · Summarize | Send outside the boundary · irreversible | ≥ MEDIUM | Privacy Officer |
| Red-Team Probe | L2 · Human-agent team | PUBLIC | Analyze | Publish externally · Send outside the boundary · Execute a tool with side effects · Write to a system of record · irreversible | ≥ MEDIUM | Security Lead |
Alice on this site runs over the synthetic Lakeview General Hospital environment. Verdicts, checks, and receipts are computed by the Nexus decision engine on synthetic requests; nothing here is a customer, a patient, or a production deployment. Signed-in organizations see their own agents, decisions, and receipts.