HomeProductsAgentic Sidecar
Agentic Sidecar · Supervise
Before the action,
check the intent.
Evaluate proposed agent actions against policy, risk, intent, and budget before they execute. Observe decisions or govern them with explicit intervention and approval boundaries.
pip install agentic-sidecar
- Open source
- Observe and Govern modes
- Explicit approval boundaries
Inspect decisions before actions execute.
Follow a sample agent plan through policy, intent, and budget checks, including a blocked action and human escalation.
- order_lookupALLOWread scope
- check_refund_policypendingpolicy check
- issue_refundpendingauthority check
- draft_responsependingintent check
Order lookup allowed: read access is within the task’s permitted scope.
Example data, simulated in your browser. No live telemetry or real actions.
Governed Customer Support Agent Demo
Test simulated agent actions against Policy Advisor, Risk Evaluator, Intent Guardian, and Budget Guardian in real time.
// Select an action to view the serialized Decision & DecisionContext payload.
Core Architecture & Design Principles
Built on explicit governance boundaries rather than implicit LLM prompts.
Rules Before Models
Policy Advisor, Risk Evaluator, Intent Guardian, and Budget Guardian run deterministically with zero LLM latency or API costs. LLM Judges are strictly opt-in.
Framework Independence
The Sidecar runtime (core/) has zero dependencies on specific framework adapters. It attaches to LangGraph, CrewAI, or custom loops seamlessly.
Fail-Safe Failure Modes
Requires an explicit on_sidecar_failure setting (fail_closed or fail_open). If evaluation errors, governance is guaranteed.
Composable Decision Gate Modules
Enable only the evaluation components your application requires.
Policy Advisor
YAML-driven allow/deny permission rules matching tool names via glob patterns (delete_*).
Risk Evaluator
Classifies actions into LOW, MEDIUM, or HIGH risk using tool argument pattern checks.
Intent Guardian
Validates tool parameters against structured user IntentEnvelope constraints and detects drift.
Budget Guardian
Enforces per-task spend (USD) and token ceilings, returning PAUSE for human approval when exceeded.
Plan Evaluator
Evaluates entire agent plans to detect unnecessary steps, overreach, or contradictory operations.
Critic & Judge Evaluators
Challenges decisions for unsupported assumptions and provides model-agnostic LLM decision review.
Observe Mode vs. Govern Mode
Gradually transition from monitoring to active policy enforcement.
| Mode | Behavior | Tool Execution | Use Case |
|---|---|---|---|
| Observe Mode | Evaluates tool calls, computes decision verdicts, and logs audit trails. | Always Executes | Initial deployment, shadow auditing, and baseline evaluation. |
| Govern Mode | ALLOW/WARN proceed. BLOCK stops; other outcomes resolve through approval or suspend execution. | Blocked via Exception | Production enforcement, financial boundary control, and safety gating. |
Integration Code Snippets
Simple, pythonic integration with LangGraph and agent frameworks.
# Wrap tools for LangGraph agent
from agentic_sidecar import Sidecar
from agentic_sidecar.adapters.langgraph import attach
from agentic_sidecar.gate.policy import PolicyAdvisor, PolicyRule
sidecar = Sidecar(
on_sidecar_failure="fail_closed",
mode="govern",
policy=PolicyAdvisor([PolicyRule(tool="delete_*", effect="deny")])
)
# Intercept tools before passing to react agent
tools = attach(sidecar, [read_order, issue_refund, delete_customer])
agent = create_react_agent(model, tools=tools)
Production Use Cases
Reproducible walkthroughs showing Agentic Sidecar securing real-world agent workflows.
Prevent Unauthorized High-Value Refunds
Intercept agent tool calls and enforce policy limits to block unauthorized monetary refunds in under 10 minutes.
View Use Case →Run a Governed Customer Support Agent
Combine Policy Advisor, Intent Guardian, Risk Evaluator, and Budget Guardrails into a unified synchronous supervision sidecar.
View Use Case →Govern LangGraph Agent Workflows
Attach Agentic Sidecar's native LangGraph adapter to intercept graph tool invocations and enforce intent constraints in govern mode.
View Use Case →Framework adapters and scoped supervision
Shared and Exclusive context and guardrails compose across native tool adapters. Governed sync/async calls share approvals, stop controls and audit evidence.
LangGraph Available
Wrap tool lists with attach(sidecar, tools) — works with create_react_agent.
AutoGen Available
Native typed tools with governed outcome handling.
CrewAI Available
Native typed tools with governed outcome handling.
OpenAI Agents SDK Available
Native typed tools with governed outcome handling.
Google ADK Available
Native typed tools with governed outcome handling.
Any Python Agent Available
Use sidecar.evaluate(context) directly in any custom agent loop.
Built for agent engineers and platform teams
AI Agent Developers
Add governance to LangGraph or any custom Python agent with a few lines of code.
Platform Safety Engineers
Define and enforce policy, risk, and budget boundaries without modifying the agent model itself.
ML / Ops Teams
Audit every agent decision with full provenance before moving from shadow-mode to production enforcement.
Start building
Put Agentic Sidecar to work.
Start with the documentation, explore the examples, or contribute on GitHub.