$npx governancebench run --target https://your-platform-url
Self-hosted by design. Bring your own model, bring your own key. Your data never leaves your infrastructure. Read the sovereignty model →
Runtime events from governed agents.
Every agent action passes through identity, permission, audit, and policy gates. The events below are sampled from a demo environment.
Governance event streamLIVE 12:42:08
Agent
Action
Policy
Result
Time
PRPermissionRouter
data_access_request
vendor-access
Blocked
12:42:08
ALAuditLogger
output_validation
audit-trail
Allowed
12:41:54
TSTrustScorer
policy_check
runtime-score
Allowed
12:41:37
RMRuntimeMonitor
behavior_drift
drift-threshold
Flagged
12:40:58
HPHaltProtocol
halt_initiated
unsafe-action
Blocked
12:40:21
ECEthicsConstraint
prompt_review
restricted-output
Flagged
12:39:46
DLDecisionLog
evidence_recorded
decision-proof
Allowed
12:39:12
Governance events today: 12,842Demo environment
Discover. Enforce. Prove. Govern.
A practical governance loop: find the agents already operating, enforce policy at runtime, produce evidence, and govern the system as it evolves.
01 Discover
Find agents your organization does not know it runs.
Map agent activity, ownership, frameworks, connectors, model routes, and runtime permissions across the stack.
02 Enforce
Policies that enforce themselves at runtime.
Turn policy into runtime decisions: allow, block, quarantine, route, or review.
03 Prove
Audit-ready proof, not audit-ready promises.
Convert policy decisions and runtime events into traceable evidence with cryptographic chain-of-custody.
04 Govern
One governance layer for every agent framework.
Coordinate governance across frameworks, connectors, models, and runtime environments.
Governance has to produce evidence.
Agentomy is designed around traceability. Every policy decision, runtime exception, connector request, and approval path becomes part of an evidence trail.
Adversarial governance testing. GovernanceBench tests whether the governance layer itself can be bypassed.
Vendor-neutral halt. One command stops all connected agents regardless of platform, with cryptographic proof of halt.
Our governance benchmark is published open source under Apache 2.0. Every claim is reproducible from a fresh clone.
Real incidents, real consequences.
Documented governance failures from the last 90 days. Each maps to specific Agentomy detection capabilities.
Lilli Breach, February 2026
46.5M messages exfiltrated in under 2 hours.
Autonomous agent exploited 22 unauthenticated API endpoints at a major consultancy. 57K user accounts, 384K AI assistants compromised. 95 system prompts were writable; attacker could reprogram thousands of consultants' AI without deploying code. No agent identity verification, no authorization checks, no audit trail, no kill switch. (CodeWall disclosure, February 2026.)
Attacker re-registered a maintainer's expired email domain, hijacked npm account, published three poisoned versions. Fires on require() with no postinstall script. First npm attack targeting AI tool configs.
Detected by: IOC Package Flagging, Version Pinning, C2 Domain Detection, Obfuscated Script Detection
Six verticals. Per-vertical incident references. Per-vertical governance patterns.
57 governance threats catalogued across six verticals. 41 critical, 16 high. Each vertical page lists its threats in severity-ranked order with its matching SDK adapter.
No third-party certifications. Self-assessed readiness mapped per control.
Procurement teams filter for the cert. We do not have it. Here is what we have instead, and how to verify each claim against the actual source.
What is mapped (self-assessed, per control)
SOC 2 readiness
EU AI Act
GDPR
HIPAA readiness
ISO 27001 aligned
PCI DSS controls
NIST SP 800-53
NIST PQC planned
SOX planned
FERPA planned
12 regulatory frameworks with 100+ controls mapped. 9 at self-assessed readiness today. The remaining 3 are control-mapped and planned. Source: compliance/ directory, traceable per control.
What we honestly do NOT claim
SOC 2 Type II report
External ISO 27001 attestation
HIPAA BAA-signed
PCI DSS attestation
FedRAMP authorization
Third-party audit
Independent pen test report
ISO 42001 mapping
SOC 2 Type II is not pursued. HIPAA requires a deployer-supplied BAA. PCI DSS infrastructure controls are provided but Agentomy itself handles no payment card data. FedRAMP needs a federal sponsor and 3PAO engagement. ISO 42001 (2023 AI Management System standard) acknowledged but not yet mapped.
When you need formal compliance verification: engage a qualified auditor against the deployed Agentomy instance and the mapped controls. Honest framing protects the deployer; ambiguous "compliant" claims do not survive audit. Read the full honest disclosure →
Discover. Enforce. Prove. Govern.
Each governance capability runs in the platform layer. The data panels below sample the operating state.
Find agents your organization does not know it runs.
Maps active agents, frameworks, permissions, and runtime state before they become ungoverned risk.
Every Agentomy primitive ships in the Agent Skills format.
AI agent platforms discover capabilities through the open Agent Skills spec. Agentomy ships 16 spec-conformant skills covering every named primitive in our vocabulary, discoverable in the agent's native namespace.
SovereignSkillRegistry
Runtime-loaded registry of every spec-conformant SKILL.md. SHA-256 contentHash per skill plus registryHash (hash-of-hashes) for tamper-detect across the full inventory. Public discovery at GET /api/skills/registry/status.
SignedSkillAttestation
Ed25519 attestation binds skill content + version to the constant Agentomy issuer. Verifier returns five specific tamper reasons. Differentiates from unsigned community skill libraries.
govern.skill() SDK
Three-gate decision that turns "loaded a skill" into "loaded a governed skill". Skill exists in registry, caller tier permits, attestation verifies. const r = await agent.skill("agent-certificate", { callerTier: "Operator" }).
22 governed integration points across AI frameworks and verticals.
Agentomy Agent ships adapters that wrap the standard call surface of each supported framework or vertical with governance: identity, permission boundary, audit, override, behavioral monitoring. One govern() call swaps in the governance layer; the rest of the framework code stays the same.
AI agent frameworks · 16
LangChain
LangGraph
AutoGen
CrewAI
Haystack
Google ADK
Microsoft Agent
Semantic Kernel
Dify
Flowise
OpenAI Agents SDK
Azure AI Foundry
AWS Bedrock
Agno
Hermes
OpenClaw
Vertical adapters · 6
RPA automation
Algorithmic trading
Medical device
Autonomous vehicle fleet
Industrial IoT
Cloud infrastructure
Source-verifiable in agentomy-agent/src/adapters/. Install via npm install agentomy-agent. Each adapter exports the same governance contract; integration is one method call (agent.govern()) per framework instance.
Open where trust begins. Commercial where governance scales.
Agentomy Agent gives builders an open governed agent framework. Agentomy builds the commercial governance layer around discovery, enforcement, evidence, and enterprise control.
PermissionRouter
AuditLogger
HaltProtocol
ExecutionTimer
AgentSandbox
DecisionLog
TraceBinding
TeamCoordinator
EthicsConstraint
TrustScorer
RuntimeMonitor
DeploymentManifest
Discover, enforce, prove.
Bring Your Own Key. Self-hosted (Tier 1) or API key (Tier 2). Your governance data stays in your infrastructure. Deploy to Azure Container Apps, Kubernetes, Docker, or air-gapped environments. Self-hosted deployment eliminates network transport exposure entirely.
