AgentPolicyPack

AgentPolicyPack lets organizations define, test, and carry AI-agent governance rules across models, frameworks, providers, and deployment environments.

AgentPolicyPack 0.2.0b1 is the first public Beta release candidate. It has documented limitations and is not production certification.

Motivation

AI-agent governance should be portable, deterministic, auditable, version-controlled, and testable in CI. AgentPolicyPack provides a vendor-neutral policy bundle format, a fail-closed evaluator, policy tests, simulation, diffing, and a Python API plus CLI.

Problem Statement

Organizations often encode agent rules inside one model provider, framework, runtime, or deployment stack. That creates lock-in and makes policy behavior hard to test. AgentPolicyPack separates governance intent from enforcement systems so the same rules can be validated, compared, and evaluated offline.

Core Capabilities

• YAML and JSON policy bundles loaded with safe parsers, bounded parsed structure, and no YAML anchors, aliases, or explicit tags.
• Strict schema models that reject unknown fields.
• Structural and semantic validation with stable finding codes.
• Safe structured condition language with bounded nesting.
• Target matching for subjects, actions, resources, tools, models, providers, and environments.
• Conflict strategies: deny_overrides, allow_overrides, first_applicable, highest_priority, and only_one_applicable.
• Effects: allow, deny, require_review, limit, redact, and log_only.
• Obligations and most-restrictive limit aggregation.
• Deterministic normalization, digests, decision IDs, and reports.
• Embedded policy tests, policy-test coverage, simulation, comparison, and conservative diffing.

Architecture

Policy bundles load into typed Pydantic models. Validation runs before evaluation. Invalid bundles fail closed as indeterminate with a deny-equivalent effective decision. The evaluator performs deterministic target matching, structured condition evaluation, conflict resolution, obligation aggregation, limit aggregation, and evidence generation. External systems remain responsible for enforcing the returned decision.

Execution consumers must gate on Decision.effective_decision or Decision.allowed, not only Decision.decision. A matching allow policy can still become effectively denied when an applicable limit is exceeded.

Installation

python -m pip install agentpolicypack

For development:

python -m pip install -e ".[dev]"

Quick Start Policy

schema_version: "1.0"
bundle:
  id: customer-support-governance
  name: Customer Support Agent Governance
  version: "1.0.0"
  namespace: example.customer_support
  default_decision: deny
  conflict_strategy: deny_overrides
policies:
  - id: allow-ticket-read
    effect: allow
    priority: 100
    targets:
      subjects:
        roles: [support_agent]
      actions: [ticket.read]
      resources:
        types: [SupportTicket]
    conditions:
      all:
        - field: resource.attributes.assigned_agent_id
          operator: equals
          value_from: subject.id

Evaluation Example

agentpolicy evaluate examples/tool_governance/policy.yaml --request examples/tool_governance/allow-search-request.yaml

Policy-Test Example

agentpolicy test examples/tool_governance/policy.yaml
agentpolicy coverage examples/tool_governance/policy.yaml

CLI Examples

agentpolicy validate examples/tool_governance/policy.yaml
agentpolicy lint examples/tool_governance/policy.yaml
agentpolicy inspect examples/tool_governance/policy.yaml
agentpolicy normalize examples/tool_governance/policy.yaml
agentpolicy digest examples/tool_governance/policy.yaml
agentpolicy simulate examples/tool_governance/policy.yaml --requests examples/tool_governance/requests.yaml
agentpolicy diff tests/fixtures/diff/policy-v1.yaml tests/fixtures/diff/policy-v2.yaml
agentpolicy compare tests/fixtures/diff/policy-v1.yaml tests/fixtures/diff/policy-v2.yaml --requests tests/fixtures/diff/requests.yaml

Python API Example

from agent_policy_pack import DecisionRequest, evaluate, load_bundle

bundle = load_bundle("examples/tool_governance/policy.yaml")
request = DecisionRequest(action="tool.call", subject={"roles": ["research_agent"]})
decision = evaluate(bundle, request)
print(decision.decision, decision.effective_decision)

Conflict Strategies

deny_overrides is the default. Any matching deny controls before review or allow. allow_overrides is supported but dangerous because it may relax denies. first_applicable uses deterministic priority-descending, ID-ascending order. highest_priority selects the highest priority and resolves equal-priority mixed effects conservatively. only_one_applicable returns indeterminate when multiple policies match.

Effects, Obligations, and Limits

Policy effects express governance intent. Decisions expose a primary outcome plus structured obligations such as audit, redact, mask, require_review, enforce_limit, notify, retain_evidence, and attach_policy_context. Custom obligations must use an extension namespace and are preserved, not executed. Limits use exact Decimal handling for costs.

Simulation and Diffing

Simulation evaluates request batches offline. Comparison evaluates two bundles against the same requests. Diffing classifies changes conservatively as security relaxation, tightening, potentially breaking, or informational.

Integration Roadmap

The package includes protocol types for future adapters. Version 0.2.0b1 works independently and offline. Forge, PrivateAIStack, ModelSwapBench, OpenOntologyLite, AIAuditLog, and AIMeter integrations are deferred unless optional adapters are installed and tested in later releases.

Security Model

AgentPolicyPack uses safe YAML loading, strict unknown-field rejection, bounded condition nesting, safe glob-style patterns, deterministic serialization, and fail-closed defaults. It never uses Python eval or exec, never imports modules named by policy files, and never follows remote URLs.

Limitations

• External systems must enforce policy decisions.
• No identity-provider integration.
• No secrets management.
• No network firewall.
• No operating-system sandbox.
• No hosted policy server.
• No graphical editor.
• No general-purpose expression language.
• No arbitrary code execution.
• No legal or regulatory certification.
• Custom obligations are preserved but not executed.
• PII obligations do not provide automatic perfect PII discovery.
• Policies cannot guarantee model behavior.
• Optional framework adapters may cover only documented integration points.
• Policy diff classification is conservative and rule-based.
• Policy-test coverage is not software-code coverage.
• Bundle format may evolve before version 1.0.

Contributing

Use the development commands in docs/development.md. Keep behavior deterministic, typed, local-first, and fail-closed.

License

Apache License 2.0.

Author

sekacorn