Autonomy levels¶

Controlled vocabulary for autonomy_profile.level values that describe agent discretion and human control boundaries.

Source fields¶

Choose the highest level of discretion the pattern normally allows, not the most optimistic implementation possibility.
Human review that occurs only for rare exceptions is different from review embedded in the normal loop.
Pair autonomy choices with matching risk and approval requirements in canonical pattern entries.

Stable id: human-directed
Summary: The agent assists with analysis or drafting, while humans retain all consequential decisions and actions.
Definition: Use for workflows where the agent organizes information, drafts outputs, or surfaces context but does not independently choose or execute material next steps.
Selection guidance: Choose this when the agent behaves as a copilot and human operators remain fully accountable for every decision and action.
Related ids: recommendation-only, approval-gated
Used by patterns:
Human-directed task orchestration
Approval packet generation
Crisis briefing evidence synthesis
Research synthesis with citation verification
Analyst copilot loop
Critical protected artifact collaboration
Critical authoritative state restoration
Critical protected-priority adaptation
Critical command-window resequencing
Critical escalation authority recommendation
Critical channel-safe state packaging

Stable id: recommendation-only
Summary: The agent can analyze options and suggest next steps, but a human decides and initiates any consequential change.
Definition: Use for workflows where the agent produces rankings, plans, or escalation advice without being allowed to commit changes on its own.
Selection guidance: Choose this when high-value support comes from prioritization or reasoning, yet execution authority stays entirely with a person.
Related ids: human-directed, approval-gated
Used by patterns:
Approval-centered collaboration
Incident root cause analysis
Critical signal corroboration triage
Governed optimization bundle retuning
Schedule adjustment and replanning
Control requirement attestation recommendation
Deal desk recommendation support
Policy-constrained escalation routing
Readiness gate disposition recommendation

Stable id: approval-gated
Summary: The agent may prepare actions or plans, but explicit human approval is required before execution proceeds.
Definition: Use for workflows where an agent can assemble an execution package, trigger-ready payload, or proposed workflow step that remains blocked until a reviewer authorizes it.
Selection guidance: Choose this when the normal operating model includes a formal approval checkpoint before any consequential action.
Related ids: recommendation-only, bounded-delegation
Used by patterns:
Browser-based form completion with approval gates
Approval-gated briefing release
Approval-gated collaborative artifact release
Evidence-gated verification for release
Approval-gated triage dispatch
Risk alert triage
Approval-gated optimization-state release
Contingency plan activation gate
Approval-gated recommendation release
Approval-gated transformation release
Batch content transformation

Stable id: bounded-delegation
Summary: The agent may act autonomously within preapproved scope, thresholds, and reversible limits.
Definition: Use for workflows where humans define guardrails in advance and the agent executes routine actions without case-by-case approval unless it hits a boundary.
Selection guidance: Choose this when there is real delegated authority, but scope, spend, targets, or reversibility constraints are intentionally tight.
Related ids: approval-gated, exception-gated-autonomy
Used by patterns:
Exception-aware task execution
Change-triggered context briefing
Shared workbench orchestration
Authoritative record reconciliation
Claimed state verification
Anomaly detection review
Adaptive threshold calibration
Calendar conflict coordination
Delegated authority option ranking
Document to structured data handoff
Normalization and enrichment

Stable id: exception-gated-autonomy
Summary: The agent runs most of the workflow autonomously and escalates only exceptions, low-confidence cases, or policy conflicts.
Definition: Use for workflows where continuous operation is expected without routine approval, while humans stay available for off-nominal conditions and sensitive edge cases.
Selection guidance: Choose this when the normal path is autonomous and human attention is reserved for exceptions rather than every case.
Related ids: bounded-delegation, autonomous-with-audit
Used by patterns:
Staged change execution with rollback holds
Explainable watchlist maintenance
Queue prioritization optimization
Authoritative change coordination refresh
Change-triggered representation refresh

Stable id: autonomous-with-audit
Summary: The agent can make and execute operational decisions independently under strong logging, controls, and retrospective oversight.
Definition: Use for workflows where the agent has broad discretion to proceed without immediate human checkpoints, and governance relies on ex ante policy plus ex post audit.
Selection guidance: Choose this only when the workflow tolerates sustained autonomy and the control environment can support detailed auditability, rollback strategy, and accountability.
Related ids: exception-gated-autonomy, bounded-delegation
Used by patterns:
Workflow hand-off and completion
Adaptive review sampling-rate tuning