Skip to content

Agent Learning Governance

Public Preview

agent_learning_gov is a source package targeting Microsoft Agent Learning 0.8.x. The integration API may change before general availability.

agent_learning_gov applies Agent Governance Toolkit controls across the Agent Learning lifecycle while preserving Agent Learning's discrete-action policy architecture.

Layer Integration capability
Episode capture Pre-action and pre-tool policy checks; durable decisions, violations, risk, and tool-use metadata
Evaluation and rewards Agent Learning scoring plus bounded governance bonuses and penalties
Offline learning Governed batches, Bayesian episode exclusion from REINFORCE, candidate action validation
Policy lineage Run-backed candidates with parent policy and training-run identifiers
Promotion Explainable evaluation suite, shadow/canary/production stages, fail-closed activation
Operations JSONL audit events and signed decision, candidate, and rollout provenance

Source

Package source: agent-governance-python/agent-learning

The package name and import name are both agent_learning_gov.

Compatibility

The implementation is validated against microsoft/agent-learning tag v0.8.0. That release introduces a unified task-policy decision surface with two routes:

Authority Selection Governance treatment
low Learned softmax policy Policy checked; requires a behavior-policy action_logprob; may train with REINFORCE
full Bayesian DecisionFrame resolution Constraints and evidence are audited; no action_logprob; scored but never used as a REINFORCE sample

Full-authority decisions use confidence-weighted Bayesian evidence, information gain, robust utility, Pareto elimination, and explicit user adjudication when required. The governance integration does not reinterpret those calculations. It enforces policy around the selected action and preserves their certificate in lineage and audit data.

Install

At the time of this review, GitHub release v0.8.0 is newer than the version reported by PyPI. For repository development, use the tagged Agent Learning source and install this package from the AGT checkout. Maintainers must apply the repository's seven-day dependency review rule before installing a newly published release:

git clone --branch v0.8.0 https://github.com/microsoft/agent-learning.git ../agent-learning
python -m pip install -e ../agent-learning
python -m pip install -e "./agent-governance-python/agent-learning"

Public API

Lifecycle

  • GovernedEpisodeCapture
  • EpisodePersistenceError
  • UnresolvedDecisionError
  • PolicyAwareRewardAdapter
  • GovernedLearningRunner
  • GovernedPolicyPromotion
  • GovernanceEvaluationPack

Audit

  • AuditEvent
  • InMemoryAuditSink
  • JsonlAuditSink

Durable governance metadata

  • GovernanceTelemetry
  • PolicyDecisionRecord
  • GovernanceViolation
  • ToolUsageRecord
  • GovernanceOutcome
  • RiskLevel

All metadata is stored under the stable agent_governance key on Agent Learning episodes, rewards, policy snapshots, and training runs.

Decision, candidate, and rollout provenance uses HMAC-SHA256. The default key is process-local; durable workflows pass the same secret-backed provenance_key to capture, learning, and promotion.

Only DecisionStatus.RESOLVED decisions can start capture. Execute the action from CaptureContext.action_id, because governance may substitute a different action and clear the original behavior-policy log probability. Use authorize_tool_call() before invoking a tool; execute its effective_action_id, then pass that one-shot authorization to record_tool_call() with the same arguments.

Promotion safety

Agent Learning 0.8.0 treats LearningStore.store_policy() as both persistence and activation. The integration therefore keeps candidates inside TrainingRun.metadata through validation and canary rollout. A deployed canary and an explicit approved violation-rate baseline are required for production. Production persists signed state and approval audit, activates the candidate, and then invokes the synchronous callback. Callback rejection restores the prior active policy. A configured audit sink and run-backed candidate storage are mandatory for production.

This prevents evaluation, blocked promotion, canary persistence, or failed production preparation from silently replacing the active policy.

Next steps