Product Controls Frameworks Docs Pricing ★ GitHub · 1,247
/ 01 · THE PROMISE
v0.1.0 · OPEN-SOURCE AGENTIC CONTROL PLANE

Governance for AI Agents.

One open-source framework to enforce safety, cost, and compliance controls on any AI agent — LangGraph, CrewAI, Bedrock, Copilot, Databricks, Azure. Not a guardrail. Not a tracer. A control plane.

View on GitHub
Trusted in production · Jindal Stainless · Hindalco · Ultratech · Databricks · AWS Bedrock
/ 02 · THE STATE OF AI AGENTS IN PRODUCTION
// Confidential · Diagnosis #24-Q2

Agents are scaling
faster than governance.

You deploy an agent on Monday. By Friday it has called the refund tool 4,200 times, leaked an API key into Slack, and cost $38K — none of which is in your dashboard.

agent.trace · refund-bot-prod · run #a82f1c
LIVE
EXECUTION LOG
  1. 00:00:01STARTagent refund-bot initialized · budget $500 · tools[4]
  2. 00:00:04TOOLlookup_order(#A-8821) → ok · $0.02
  3. 00:00:07TOOLcheck_policy(refund) → within_limit · $0.02
  4. 00:00:11TOOLissue_refund($48) → confirmed · $0.04
  5. 00:00:13LOOPmodel called lookup_order again with same args
  6. 00:00:15LOOP3rd identical call to lookup_order · no circuit breaker
  7. 00:00:22DRIFTresponse tone off-policy (+0.4σ) · no gate
  8. 00:01:04LEAKassistant replied with raw STRIPE_KEY · no content scan
  9. 00:03:17COST1,412 tool calls · budget 283% over · no kill switch
  10. 00:05:48POLICYrefund granted $1,800 above tier · no approval gate
  11. 00:11:22AUDITcontract signed without signed trace · SOC2 finding
  12. --:--:--agent still running
SourceWarden incidents · n=412 · Q1 2026
Agents with ≥1 uncaught failure73%
Median days to discovery17
0%
of enterprise agents ship with at least one uncaught failure class in the first 90 days.
Gartner · Agentic AI Observer 2026
days
until the EU AI Act comes into force · €35M or 7% revenue in penalties.
Regulation (EU) 2024/1689 · 2 Aug 2026
0
frameworks today cover safety + cost + drift + audit + red-team in one layer. You stitch together 6.
Guardrails · NeMo · LangSmith · Bedrock · Lakera · glue
/ 03 · THE ANSWER
// Introducing Warden
24 Controls.
6 Frameworks.
1 Line of code.

Everything regulated enterprises need to ship agents safely — in one open-source library. Drop it into any framework. Every tool call governed by default.

Keep scrolling
/ 04 · THE ONE LINE
// From zero to governed

One function call.
Every safeguard on.

agent.py
# Drop into any existing agent. No framework rewrite.from acp import governfrom langgraph.graph import StateGraph graph = StateGraph(...)agent = govern(graph.compile(), "procurement-agent", budget=1.0) # That's it. All 24 controls now active:# ✓ Tool call interception    ✓ Budget enforcement (80/100/200)# ✓ Policy engine (30 rules)  ✓ Content scanner (34 patterns)# ✓ Drift + bias detection    ✓ Red team + approvals# ✓ BU isolation + consent    ✓ Dashboard reporting
Auto-detects framework · Wraps all tools Dashboard live
/ 05 · WHAT 24 CONTROLS ACTUALLY MEANS
// Governance stack

Six categories.
Twenty-four controls.

Catalog · 2026.Q1v0.1.0 · Stable
/ 01
Safety & Control

Keep agents bounded.

  • Loop prevention
  • Iteration caps
  • Circuit breakers
  • Autonomy tiers
  • Deadlock detection
  • Dead-man switch
04CTRL
/ 02
Cost & Ops

Stop runaway bills.

  • 3-tier budget · 80/100/200
  • Runtime blocking
  • Performance SLAs
  • Error budgets
  • Graceful degradation
02CTRL
/ 03
Policy & Compliance

Ship audit-ready.

  • NL policy engine
  • 30 pre-loaded rules
  • EU AI Act
  • GDPR erasure
  • OWASP LLM Top 10
04CTRL
/ 04
Observability & Learning

See every decision.

  • Real-time tracing
  • Langfuse + OTel
  • Decision quality
  • Human feedback
  • CoT explainability
04CTRL
/ 05
Security & Data

Zero credential leak.

  • BU-scoped mTLS
  • X.509 agent identity
  • 34-pattern scanner
  • Intent signing
  • Purview DLP
05CTRL
/ 06
Quality & Evolution

Agents that improve safely.

  • Drift detection
  • Bias checks
  • Canary rollouts
  • DSPy + RLHF
  • Hard safety boundaries
05CTRL
24
Total controls
34
Content patterns
30
Pre-loaded policies
6
Framework adapters
/ 06 · DROP INTO ANYTHING
// Native framework adapters

You already picked
an agent framework.

6 / 6 adapters100% coverage · Nov 2025
01 LangGraph
v0.2+
LangGraph
guarded_node · recursion_limit · Pregel interception
02 CrewAI
v0.70+
CrewAI
Task + Agent wrapping · tool governance
03 Azure AI
Foundry
Azure AI Foundry
Prompt Flow · Purview DLP bridge
04 AWS
AgentCore
Bedrock AgentCore
microVM isolation · Gateway interceptors
05 Copilot
Studio
Copilot Studio
Topic + Action wrapping · connector scoping
06 Mosaic AI
+ Unity
Databricks Mosaic
Unity Catalog · MLflow · Inference Tables
/ 07 · HOW GOVERNANCE FLOWS
// Architecture

A governance layer
between your agent and the world.

LAYER 01
Assessment Dashboard
19-page React + Ant Design console. Sessions · policies · approvals · costs · drift · red team results.
UI
LAYER 02
MCP-Native Gateway
FastAPI proxy. Every tool call intercepted. Works with any MCP tool server without framework changes.
Gateway
LAYER 03
Evolution Engine
DSPy optimization · evolveRL · RLHF. Agents self-improve within hard safety boundaries.
RL
LAYER 04
Core SDK — 24 Control Modules
Pre/post tool-call pipeline. Budget · policy · content scan · drift · bias · consent · audit. Defense in depth.
SDK
LAYER 05
Framework Adapters
LangGraph · CrewAI · Azure · Bedrock · Copilot · Databricks — native hooks per platform.
Adapters
LAYER 06
Observability Plane
Langfuse + OpenTelemetry + Grafana. Every decision traced, every override logged, every cost attributed.
Telemetry
/ 08 · PROOF IN PRODUCTION
// Case study № 01 · Filed FY25

19 nodes.
One line of governance.

Jindal Stainless
ABG Conglomerate · Enterprise · Live FY25
0
Governed nodes
in one import
0%
Tool-call coverage
4 + 4
Bedrock · Databricks live
0
Lines of graph code touched
"
Warden wrapped our entire NL2SPARQL agent graph with one call. Every tool invocation now has budget checks, content scanning, and audit trails — without touching our graph code.
Lead AI Architect
Jindal Stainless · Data & AI Platform
LIVE · nl2sparql.pipeline session #a8f3
  1. PASS▸ intent.parsepolicy.check · 0.3ms
  2. PASS▸ entity.resolvebudget.deduct · $0.0012
  3. SCAN▸ sparql.generatecontent.scan · 34 patterns · clean
  4. PASS▸ schema.validatemTLS · purview.allow
  5. HOLD▸ query.executeautonomy.tier=supervised · human.approval
  6. PASS▸ result.formataudit.sign · X.509
LATENCYp95 · 184ms
OVERHEAD+3.2%
BLOCKS (30d)47
BEFORE
Raw agent graph. No budget caps. No audit. No consent record. Every tool call a potential incident.
govern(graph)
AFTER
Same graph. 24 controls active. EU AI Act ready. Drift + bias monitored. Audit trail to every decision.
/ 09 · WHY NOT THE OTHERS
// Honest comparison

Every other tool
covers a slice. Warden covers the surface.

Warden Guardrails AI NeMo Guardrails LangSmith Bedrock Guardrails Lakera
24 integrated controls
Runtime budget enforcementtracking only
Self-evolving agents
MCP-native governance proxy
6 framework adaptersgenericgenericLangChain onlyBedrock onlygeneric
Autonomy tiers with approvals
Red teaming + CVSS scoringbasic
Business-unit isolation
EU AI Act · GDPR mappingpartial
Open-source · self-host
Warden100%
Guardrails AI35%
NeMo Guardrails30%
LangSmith25%
Bedrock Guardrails20%
Lakera28%
/ 10 · COMPLIANCE READY
// Mapped to standards

Every control tagged
to the regulation that demands it.

EU AI Act · Annex III · High-Risk
GDPR · Art. 17 · 22 · 35
OWASP LLM Top 10 · 2025
NIST AI RMF · Govern · Map · Measure · Manage
SOC 2 Type II · Ready
ISO 42001 · AI Management
/ 11 · START TODAY
EU AI ACTGPAI obligations · Aug 2026 COLORADO AI ACTAutomated decisions · Feb 2026 NYC LOCAL LAW 144Bias audits · active CALIFORNIA SB 1047Frontier models · pending UK AI OPPORTUNITIESAction plan · 2026 ISO 42001AI management · certifiable EU AI ACTGPAI obligations · Aug 2026 COLORADO AI ACTAutomated decisions · Feb 2026 NYC LOCAL LAW 144Bias audits · active CALIFORNIA SB 1047Frontier models · pending UK AI OPPORTUNITIESAction plan · 2026 ISO 42001AI management · certifiable
// Nothing to lose. Billions to govern.

Ship governed agents before the regulators ship you.

Install the SDK. Wrap your agent. See every tool call in your dashboard within sixty seconds.

01
t + 0s
pip install
02
t + 20s
govern(graph)
03
t + 60s
traces live
The other path
  • Wait for an incident
  • Retrofit governance under regulator pressure
  • Pay the 7% GPAI fine
  • Explain the autonomy tier to your board
OR
This path
  • One import
  • 24 controls active in 60 seconds
  • Audit-ready by lunch
  • Sleep through the Aug 2026 deadline
MIT Licensed· Self-hosted or Cloud· RSA-4096 Licensing· Full GDPR & CCPA