abi-policy

Deterministic policy evaluation engine for AI agent governance. Evaluate routing rules, tool allowlists, and budget thresholds against structured decision requests.

Zero required dependencies.

Install (after PyPI publication)

pip install abi-policy

Status: Not yet published to PyPI. Install from source: pip install -e . from the repo root.

Quick start

Policy evaluation

from abi_policy.engine import PolicyEngine
from abi_policy.models import DecisionRequest

# Load a policy bundle with tool allow/deny lists and escalation rules
bundle = {
    "bundle_id": "default",
    "rules": {
        "tools": {"allow": ["web_search", "file_read"], "deny": ["rm_rf"]},
        "escalation": {"risky_actions": {"override:*": "require_approval"}},
    },
}

engine = PolicyEngine(bundle)

# Evaluate a decision
request = DecisionRequest(action="tool_call", resource="web_search", run_id="run-42")
response = engine.evaluate(request)

print(response.decision)    # "allow" | "deny" | "require_approval"
print(response.reason_code) # e.g., "TOOL_ALLOWED"
print(response.explanation) # human-readable explanation

Capability matrix

from abi_policy.capability_matrix import CapabilityMatrix, AgentProfile

# Define agent capabilities with per-task confidence scores
profile = AgentProfile(
    agent_id="claude_code",
    agent_version="claude-sonnet-4",
    confidence_by_task={"code_write": 0.9, "test_run": 0.85},
    failure_patterns=["timeout on large repos"],
)

matrix = CapabilityMatrix()
matrix.add_profile(profile)
best = matrix.get_best_agent("code_write")  # "claude_code"

What's included

• PolicyEngine -- deterministic rule evaluation against structured DecisionRequest/DecisionResponse models
• Capability matrix -- agent profile management with capability boundaries
• Boundary registry -- track and enforce cross-component access boundaries
• Escalation rules -- pattern-matching for risky actions that require approval
• Tool allowlists -- declarative control over which tools an agent can invoke

Use it standalone

abi-policy works on its own for any AI agent governance scenario. Define policy bundles, load them into the engine, and evaluate decision requests. No ABI ecosystem integration needed.

Within the ABI ecosystem, it integrates with the orchestrator for pre-flight agent decisions.

ABI ecosystem

Related packages:

• abi-control-core -- contracts and audit trail SDK
• abi-evals -- eval runner and golden management
• abi-observability -- OTLP-compliant telemetry

Versioning

Follows Semantic Versioning. Current version: 0.1.2. See CHANGELOG.md for release notes.

License

MIT License (see LICENSE file).