autonomy-complexity.md

title	Complexity-proportional autonomy
description	Declare and score intent and decision complexity so agent-learning requires proportionally stronger evidence before autonomous execution.
author	Microsoft
ms.date	2026-08-10
ms.topic	concept
keywords	agent autonomy decision complexity intent complexity evidence thresholds human approval
estimated_reading_time	10

Complexity-proportional autonomy

A single autonomy threshold is too permissive for consequential decisions and too expensive for routine ones. agent-learning therefore separates two questions:

1. How complex is this recurring intent and decision?
2. Has the policy collected enough evidence for that complexity?

Complexity changes the required evidence. It never changes observed rewards, correctness labels, policy probabilities, or deterministic approval rules.

Declare complexity; do not infer it from prose

The agent must not grade the complexity of its own answer from free text. That would make autonomy sensitive to wording and create an incentive to understate risk. Instead, each decision policy has a validated, persisted complexity profile. The number of actions is derived by the SDK.

{
  "intent_ambiguity": "medium",
  "context_variability": "variable",
  "outcome_observability": "delayed",
  "decision_impact": "high",
  "reversibility": "costly",
  "requires_human_approval": false,
  "rationale": "Architecture choice affects a shared production integration."
}

Field	Values	Meaning
intent_ambiguity	low, medium, high	How much interpretation is required before alternatives are comparable.
context_variability	stable, variable, dynamic	How often material conditions change across executions.
outcome_observability	direct, delayed, subjective	How independently and quickly success can be measured.
decision_impact	low, medium, high, critical	The scope of cost, quality, safety, or downstream consequences.
reversibility	reversible, costly, irreversible	The cost or feasibility of undoing the selected action.
requires_human_approval	Boolean	A deterministic override. When true, learned autonomy is never granted.
rationale	String	Human-readable justification for review and audit.

Missing profiles use a conservative standard default. New integrations should always declare a profile. Existing policies can be configured without creating a learned snapshot; changing configuration must not fabricate policy stability.

Complexity score

Intent complexity is the sum of three ordinal values:

Dimension	0	1	2
Ambiguity	low	medium	high
Variability	stable	variable	dynamic
Observability	direct	delayed	subjective

Decision complexity adds:

Dimension	0	1	2	3
Impact	low	medium	high	critical
Reversibility	reversible	costly	irreversible	—
Action-space size	2 actions	3–4 actions	5+ actions	—

The total ranges from 0 to 13:

Total	Tier
0–2	low
3–6	standard
7–9	high
10–13	critical

Risk floors prevent averaging away a severe dimension:

• critical impact always produces the critical tier;
• high-impact irreversible decisions are critical;
• high ambiguity plus subjective outcomes are at least high.

Proportional evidence

Every statistical criterion remains conjunctive. A policy earns statistical autonomy only when all criteria for its tier pass.

Criterion	Low	Standard	High	Critical
Scored outcomes	3	20	50	100
95% Wilson correctness lower bound	40%	90%	95%	97.5%
Mean aggregate reward	> 0.00	> 0.00	> 0.10	> 0.20
Winner probability	Not gated	60%	75%	85%
Margin over runner-up	Not gated	15 points	30 points	45 points
Consecutive trained snapshots with same winner	1	3	4	5
Autonomous drift-audit rate	5%	10%	25%	50%

The low tier is calibrated for bounded, reversible preferences with direct feedback: three unanimous, positively rewarded outcomes can earn autonomy after one policy update, while mixed correctness evidence does not. Higher tiers keep strong probability and stability gates. Their Wilson thresholds usually require more correctness labels than the scored-outcome minimum. Global environment variables can raise or lower numeric thresholds, but the resolved values are always returned in autonomy.criteria for audit.

Explicit user acceptance

Explicit acceptance is authorization, not a statistical sample. A completed episode with metadata.feedback_status: accepted grants autonomy immediately for that (agent_id, task_id) decision and pins its accepted action_id, even when the policy logits currently prefer another action. The assessment returns authorization_basis: user_acceptance, sets the audit rate to zero, and does not request routine or drift-audit feedback again.

The newest explicit feedback wins. A later completed episode with metadata.feedback_status: rejected revokes the acceptance and returns the policy to its statistical evidence gates. requires_human_approval: true continues to block autonomy because it represents a deterministic approval requirement rather than a reusable policy preference.

Human approval and execution authority

requires_human_approval: true adds a failing autonomy criterion regardless of probability or outcome history. The SDK reports the block; it never silently converts approval into a statistical threshold.

The profile does not grant operating-system, cloud, financial, clinical, security, or destructive-action permission. Existing deterministic approvals remain authoritative even when the policy is statistically autonomous.

Lifecycle

1. Initialize a policy with --complexity-profile, or configure an existing policy with task-policy-complexity-set.
2. task-policy-decide resolves the profile, derived action-space points, tier, and proportional threshold set.
3. The response exposes autonomy.complexity, every threshold, every observed value, and the resulting mode.
4. Statistically autonomous execution continues recording observable outcomes and requesting tier-proportional drift audits. Explicitly accepted policies continue recording outcomes without requesting feedback.
5. Negative outcomes, a changed winner, or a stricter profile can immediately return the next decision to supervised mode.

Complexity profile changes update policy configuration in place. They do not increment policy version or count as a stable learned snapshot.