Skip to content
AletheIA Docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

Execution Pattern Selection — Specification

Purpose

Define the per-task pattern selection declaration: before any execution starts, a task must declare its shape, its risk, the recommended execution vehicle, the recommended execution pattern, the controls the selection requires, and an explicit selection verdict. The conceptual background — why topology is selected, and how it differs from depth and authority — is in concepts/execution-pattern-governance.md.

This declaration is filled before execution and sits upstream of the Agent Harness Contract (ADR-013): first the topology of execution is selected, then the per-task envelope (autonomy, tools, gates) is declared, then the harness (AHGE) authorizes per action. Selection must happen before the skill executes.

Non-goals

  • No runtime, scheduler, router, or policy engine — this is a declaration to be reviewed, not a mechanism that executes.
  • No restating of AHC envelope fields, AHGE permission logic, or REGC effort logic — the selection references them; the layers stay separate.
  • No new autonomy or verdict vocabulary beyond the four selection verdicts declared here. Authority is declared with autonomy levels; enforcement outcomes use the policy verdicts.

Declaration shape

execution_pattern_selection:
  task_id:
  task:
  work_type: engineering | product | research | design | governance | communication | operations

  task_shape:
    repeated: true | false
    decomposable: true | false
    independent_units: true | false
    stages_depend_on_previous_output: true | false
    requires_judgment: low | medium | high
    objective_verification_available: true | false
    state_required: true | false

  risk:
    risk_level: low | medium | high | critical
    touches_sensitive_context: true | false
    external_side_effect: true | false
    irreversible_action_possible: true | false

  recommended_vehicle:
    type: manual_prompt | single_agent | orchestrated_workflow | loop | human_led_workflow
    rationale:

  recommended_pattern:
    type: manual_prompt | single_agent | classify_and_act | fan_out_and_synthesize | adversarial_verification | generate_and_filter | tournament_compare | loop_until_done | scheduled_stateful_loop | human_led_workflow
    rationale:

  required_controls:
    state_required: true | false
    objective_gate_required: true | false
    maker_checker_required: true | false
    human_review_required: true | false
    token_budget_required: true | false
    audit_record_required: true | false

  decision:
    verdict: approved | approved_with_constraints | rejected | human_led_required
    notes:

Normative rules

  1. Selection precedes execution. The declaration must exist and carry a verdict before the task’s Agent Harness Contract is filled and before any skill runs.
  2. A rationale is required. Both recommended_vehicle.rationale and recommended_pattern.rationale must state why this topology is proportional to the task shape and risk — proportionality, not capability, justifies the choice.
  3. Loop patterns inherit the loop rule. If recommended_pattern.type is loop_until_done or scheduled_stateful_loop, the selection must satisfy the Objective Gate Policy: objective stop condition, budget, persistent state when recurring, objective gate when artifacts change, and human review before irreversible action. Without an objective gate, the verdict must be rejected or human_led_required.
  4. No verification, no autonomy. If task_shape.objective_verification_available is false, autonomous loop vehicles are not admissible; the vehicle degrades toward human_led_workflow (see Execution Vehicle Selection).
  5. Irreversibility requires human review. If risk.irreversible_action_possible is true, required_controls.human_review_required must be true.
  6. The verdict is a selection verdict, not a permission verdict. approved | approved_with_constraints | rejected | human_led_required decides whether this topology may be used. Per-action permission outcomes remain the AHGE harness decision values and the policy verdicts projection — this declaration does not replace them.

Relationship to the layer chain

request → task assessment → execution pattern selection (this declaration)
  → orchestration contract (when orchestrated)
  → agent harness contract (per-task envelope)
  → skills → harness authorization (AHGE) → gates → audit → human review

The selection consumes the REGC effort and risk classification (runtime-effort-governance-contract.md); it does not re-decide effort. Each selected topology executes under an Agent Harness Contract; per-action authorization and the governance record stay with the Agent Harness Governance Extension.

Was this page helpful?