Agent Control

Core Positioning#

Agent Control does not participate in Agent business logic orchestration; it serves as a governance layer (Control Plane) independent of the Agent. It inserts interceptors at key nodes—input, LLM output, tool calls, and final actions—and performs real-time validation and intervention based on predefined policies.

Policy & Governance#

• Centralized Control: Define a control rule once, apply it across all Agents; policy updates do not require redeploying Agent applications.
• Composable Condition Trees: Build complex composite validation conditions using and, or, not logical operators.

Runtime Interception & Decision#

• Lifecycle Coverage: Real-time validation at pre (input) and post (output) stages via the @control() decorator.
• Multi-dimensional Decision Control: Four control actions: allow, deny (throws ControlViolationError), steer (guide/rewrite), warn (warn and pass).

Evaluation Engine#

• Built-in Evaluators: Regex, List, JSON, SQL evaluators available out of the box.
• Pluggable Extensions: Support for custom evaluator integration.
• Third-party Guardrail Integration: Native support for Amazon Bedrock Guardrails, NVIDIA NeMo Guardrails, Galileo Luna-2, Azure AI Content Safety, Cisco AI Defense.

Execution & Performance#

• Dual Execution Modes: server (remote server-side evaluation) and local (local client-side evaluation).
• Client-side Caching: Policy caching in local mode to reduce network overhead.
• High-performance Benchmarks: Single Control evaluation reaches 437 RPS (p50 36ms); 50 parallel Control evaluations reach 199 RPS (p50 63ms).

Observability#

• Audit Logs: Complete recording of Agent workflow trajectories and Control trigger events.
• Custom Telemetry: Support for registering custom ControlEventSink to forward events to external logging or monitoring systems.
• Visual Dashboard: Built-in Web Dashboard providing Agent registration and Control visual management.

Architecture Overview#

Core modules: Engine (evaluation engine), Evaluators (evaluator pool), Control Store (policy storage backed by PostgreSQL), Server (API/gRPC service), Telemetry (telemetry module), Models (data models), UI (console). Directory structure: engine/, evaluators/, server/, sdks/ (multi-language SDKs), models/, telemetry/, ui/, docs/, examples/.

Typical Use Cases#

Framework Compatibility#

Out-of-the-box support for LangChain / LangGraph, CrewAI, Google ADK, AWS Strands, OpenAI Agents SDK, AutoGen, and more.

Quick Start#

Prerequisites: Docker, Python 3.12+

# One-click launch (PostgreSQL + Agent Control Server + UI)
curl -L https://raw.githubusercontent.com/agentcontrol/agent-control/refs/heads/main/docker-compose.yml \
  | docker compose -f - up -d

uv venv && source .venv/bin/activate
uv pip install agent-control-sdk

import agent_control
from agent_control import control, ControlViolationError

@control()
async def chat(message: str) -> str:
    return await LLM.ainvoke(message)

agent_control.init(agent_name="my_bot", agent_description="My Chatbot")

Unconfirmed Information#

• Initial release date not explicitly stated in README or official site
• Enterprise/paid edition page not clearly found
• PyPI package version and TypeScript SDK npm package name need repository confirmation
• OIDC / SAML enterprise identity authentication support not mentioned in docs
• Native multi-tenancy isolation not explicitly stated in README