Agency control loop

The compounding AI skill loop

A visual projection of existing AgentLoopDefinition semantics: bounded intention, state, typed action, independent verification, safety, receipt, and a closed learning return.

The picture and ordered text below convey the same control loop.

Picture first

The compounding AI skill loop

pipeline-exemplar

A bounded intention becomes typed action, independent verification, a proof receipt, and a sharper next setpoint through the existing AgentLoopDefinition control loop.

Source: libs/agency/src/lib/pipeline/control-loop/loop-contract.ts#AgentLoopDefinition

AIL-G01Trigger + bounded intention[FC]
HAHuman outcome owner[L3]●
human-gatemedium risk

The outcome owner defines what may change and the stop boundary.

appliesTo: Trigger + bounded intention
intent
AIL-M02Load prior state + evidence
OALoop orchestrator[L2]●
compute

statePolicy loads caller-managed or durable prior state before action.

appliesTo: Load prior state + evidence
context
AIL-C03Perform typed skill pipeline
DASkill agent (maker)[L1]●
ai

The maker performs the bounded action through typed pipeline steps.

appliesTo: Perform typed skill pipeline
candidate + evidence
AIL-V04Independent verification
DAIndependent verifier (checker)[L2]●
aiMAKER-CHECKER

Maker and checker identities cannot collapse when separation is required.

appliesTo: Independent verification
verdict
AIL-G05Cost · isolation · escalation gates[FC]
OALoop orchestrator[L2]●
branchcritical risk

Budget, isolation, failed verification, and escalation policy stop unsafe continuation.

appliesTo: Cost · isolation · escalation gates
[safe]accepted outcome
AIL-R06Record proof receipt
OALoop orchestrator[L2]●
compute

The receipt records outcome, evidence, cost, verification, and stop reason.

appliesTo: Record proof receipt
receipt
AIL-T07Variance + nextSetpoint
HAHuman outcome owner[L3]●
compute

VVFL learning sharpens, narrows, or stops the next pass.

appliesTo: Variance + nextSetpoint

Feedback controls

variance-next-setpoint
[continue]VVFL learning
bounded-intention
appliesTo: VVFL learning
  • The human owns the outcome and escalation boundary.
  • The maker acts; an independent checker verifies.
  • Unsafe continuation stops before a receipt is accepted.
  • Variance and nextSetpoint close the loop and sharpen the next pass.

Equivalent text

Follow the same flow without relying on position, shape, colour, or connector direction.

  1. 01

    Trigger and bound the intention

    The human outcome owner defines the desired change, authority, and stop boundary.

  2. 02

    Load prior state and evidence

    The loop orchestrator uses statePolicy before work starts, so the pass begins with explicit context.

  3. 03

    Perform the typed skill pipeline

    The skill agent acts as maker through typed steps; it does not silently grade itself.

  4. 04

    Verify independently

    A distinct checker evaluates the candidate and its evidence against the verification policy.

  5. 05

    Apply safety gates

    Cost, isolation, verification, and escalation policies stop unsafe continuation.

  6. 06

    Record a proof receipt

    The loop records outcome, evidence, cost, verification, and the reason it continued or stopped.

  7. 07

    Return learning through VVFL

    Variance and nextSetpoint sharpen, narrow, or stop the next bounded pass.