strands-agt

Microsoft Agent Governance Toolkit (AGT) policy enforcement for Strands Agents.

Provides a deterministic, YAML-based InterventionHandler that gates tool calls before execution — no model judgment involved.

Installation

pip install strands-agt

Quick Start

from strands import Agent
from strands_agt import AGTGovernance

governance = AGTGovernance(policies="policies.yaml")
agent = Agent(tools=[...], interventions=[governance])
result = agent("search for documents")

Policy Format

Policies are defined in YAML:

policies:
  - name: allow-search
    effect: allow
    actions: ["search"]
    principals: ["User::*"]

  - name: deny-delete
    effect: deny
    actions: ["delete_file"]
    reason: "File deletion is not permitted"

Effects

Evaluation Order

1. Deny policies are checked first — first match short-circuits
2. Steer policies are checked next
3. Allow policies are checked last — first match permits
4. No match = deny (fail-closed)

Features

Role-Based Access

governance = AGTGovernance(
    policies="role_policies.yaml",
    principal_resolver=lambda state: {
        "type": state.get("user_role", "User"),
        "id": state.get("user_id", "anonymous"),
    },
)

agent = Agent(tools=[...], interventions=[governance])
agent("Delete backups", invocation_state={"user_role": "Admin", "user_id": "alice"})

Rate Limiting

policies:
  - name: rate-limited-email
    effect: allow
    actions: ["send_email"]
    conditions:
      session.call_count:
        lt: 5

Context Enrichment

governance = AGTGovernance(
    policies="env_policies.yaml",
    context_enricher=lambda ctx: {
        "environment": os.environ.get("DEPLOY_ENV", "development"),
        "tenant_id": ctx["invocation_state"].get("tenant_id"),
    },
)

Condition Operators

Conditions support: lt, lte, gt, gte, eq, neq, in.

conditions:
  session.call_count:
    lt: 10
  input.database:
    in: ["analytics", "staging"]

Configuration

Development

pip install -e ".[dev]"
pytest
ruff check src tests
mypy src

References

• Microsoft AGT
• Strands Agents
• Strands Interventions docs