Closing the Loop: Autonomous Quantum Platform Operations

The long-term operating goal is not a fully autonomous quantum computer that hides all complexity. That would be unsafe and misleading. The goal is supervised autonomy: systems that can observe, diagnose, recommend, and sometimes act within explicit guardrails.

Autonomous operation is a control problem across the whole platform: calibration, compilation, scheduling, mitigation, routing, evidence packaging, cost control, and user guidance. Each action must have a policy boundary and a rollback path.

Diagram loads as you read

Closing the Loop: Autonomous Quantum Platform Operations · Figure 1

View diagram source

flowchart LR
    Observe[Observe] --> Diagnose[Diagnose]
    Diagnose --> Plan[Plan]
    Plan --> Act[Act]
    Act --> Verify[Verify]
    Verify --> Learn[Learn]
    Learn --> Observe

Autonomy levels

Diagram loads as you read

Autonomy levels · Figure 2

View diagram source

flowchart TB
    L0[L0 manual] --> L1[L1 recommend]
    L1 --> L2[L2 approve-to-act]
    L2 --> L3[L3 bounded auto-action]
    L3 --> L4[L4 adaptive policy]

Level	Description	Example
L0	humans inspect and act	manual recalibration decision
L1	system recommends	suggest target drain
L2	human approves action	approve reroute or recalibration
L3	bounded auto-action	auto-reject production job on red health
L4	adaptive policy	adjust benchmark cadence within approved limits

Do not jump to L4 for safety-critical or claim-critical actions.

Control domains

Diagram loads as you read

Control domains · Figure 3

View diagram source

flowchart TB
    Autonomy[Autonomy domains] --> Calibration[Calibration]
    Autonomy --> Scheduling[Scheduling]
    Autonomy --> Compilation[Compilation]
    Autonomy --> Mitigation[Error mitigation]
    Autonomy --> Evidence[Evidence packaging]
    Autonomy --> Cost[Cost controls]
    Autonomy --> Support[User support]

Each domain needs a separate risk model.

Guardrail architecture

Diagram loads as you read

Guardrail architecture · Figure 4

View diagram source

sequenceDiagram
    participant Agent
    participant Policy
    participant Simulator
    participant Human
    participant Runtime

    Agent->>Policy: propose action
    Policy-->>Agent: allowed with conditions
    Agent->>Simulator: dry-run if required
    Simulator-->>Human: evidence summary
    Human-->>Runtime: approve high-risk action
    Runtime-->>Agent: result and audit record

Guardrails should be machine-enforced, not policy documents in a shared drive.

Action taxonomy

Diagram loads as you read

Action taxonomy · Figure 5

View diagram source

flowchart TB
    Action[Autonomous action] --> Safe[Safe reversible]
    Action --> Reversible[Reversible but impactful]
    Action --> Risky[High-risk]
    Action --> Forbidden[Forbidden]

Examples:

Class	Example	Control
safe reversible	add extra validation probe	budget limit
reversible but impactful	reroute nonproduction jobs	audit log
high-risk	change calibration policy	human approval
forbidden	modify evidence after claim approval	block

Autonomy evidence

Diagram loads as you read

Autonomy evidence · Figure 6

View diagram source

flowchart LR
    Proposal[Action proposal] --> Reason[Reason]
    Reason --> Policy[Policy decision]
    Policy --> Execution[Execution record]
    Execution --> Outcome[Outcome]
    Outcome --> Learning[Learning update]

Every autonomous action should leave an evidence trail:

Illustrative listing · yaml

autonomous_action:
  action_id: auto_2026_04_19_001
  proposed_by: calibration_agent_v4
  reason: two_qubit_error_trend_exceeded_policy
  policy_version: qpu_health_policy_7
  risk_class: reversible_but_impactful
  dry_run_required: true
  approval: human_required
  outcome: target_restricted_for_production
  rollback: restore_prior_admission_policy

Learning loops

Diagram loads as you read

Learning loops · Figure 7

View diagram source

flowchart TB
    History[Operational history] --> Features[Feature store]
    Features --> Model[Recommendation model]
    Model --> Recommendation[Recommendation]
    Recommendation --> HumanReview[Human review]
    HumanReview --> Outcome[Outcome]
    Outcome --> History

The learning loop should learn from rejected recommendations, not only accepted actions.

Failure modes of autonomy

Diagram loads as you read

Failure modes of autonomy · Figure 8

View diagram source

flowchart TB
    Failure[Autonomy failure] --> OverAct[Over-action]
    Failure --> UnderAct[Under-action]
    Failure --> Feedback[Bad feedback loop]
    Failure --> Objective[Wrong objective]
    Failure --> Opacity[Unexplainable action]
    Failure --> Drift[Model drift]

Mitigations:

Failure	Mitigation
over-action	action budgets and cooldowns
under-action	missed-detection review
bad feedback loop	holdout probes and causal review
wrong objective	multi-objective policy including claim risk
opacity	explanation requirement
model drift	model monitoring and periodic reset

Human operating cadence

Diagram loads as you read

Human operating cadence · Figure 9

View diagram source

gantt
    title Supervised autonomy cadence
    dateFormat  YYYY-MM-DD
    section Daily
    Autonomous action review :a1, 2026-04-20, 1d
    section Weekly
    Policy threshold review :b1, 2026-04-21, 1d
    section Monthly
    Model performance review :c1, 2026-05-01, 2d
    section Quarterly
    Autonomy risk review :d1, 2026-07-01, 3d

Autonomy does not remove humans. It changes their job from routine reaction to policy design, exception review, and system learning.

Roadmap for adoption

Diagram loads as you read

Roadmap for adoption · Figure 10

View diagram source

flowchart LR
    Instrument[Instrument everything] --> Recommend[Recommendation-only mode]
    Recommend --> Approve[Approve-to-act mode]
    Approve --> Bounded[Bounded auto-action]
    Bounded --> Audit[Audit and expand]

Start with observability. A platform that cannot explain its past should not automate its future.

Operating rule

The safest autonomous quantum platform is not the one that acts most often. It is the one whose actions are bounded, explainable, reversible, evidenced, and aligned with the scientific claim being made.