refund.createRefunds and credits
Require approval for high-value refunds. Deny suspicious refund loops. Preserve proof for audit.
Control what your AI agents can do.
Protect production actions with runtime policies, human approvals, deterministic halts and Proof Packs.
Not sure where to begin? Run a 3-minute action-risk triage
Intercepttool calls before execution
Enforceallow, deny, approve, quarantine
Verify signed evidencewith record-level integrity and assurance metadata
Redactcustomer data stays protected
From first risky action to active runtime control.
HaltState starts with one concrete side effect, then guides the operator from SDK install through draft policy review and runtime evidence.
1Identify one risky action
2Install the SDK
3Let the Policy Assistant draft controls
4Test ALLOW, APPROVAL_REQUIRED and DENY
5Activate when ready
After installation, the onboarding wizard can detect or register the agent/action, guide risk selection, let the AI Policy Assistant draft policies, keep a human review step before publication, enforce at runtime, and preserve Proof Pack evidence. The public triage tool prepares a starter action map; it does not activate production policy.
Govern the actions that create real business risk.
Retail and customer operations agents can do real good and real harm. HaltState governs the business actions that affect your customers, your money, and your data.
payment.authorizePayments and purchasing
Enforce thresholds, pause unusual spend, and quarantine runaway agents before money moves.
customer.pii.exportCustomer data access
Block unnecessary PII exports, require approval for sensitive reads, and redact evidence safely.
customer.email.sendCustomer communications
Review outbound emails and SMS, stop policy-violating messages, and keep generation-labelled decision evidence.
database.writeProduction and database writes
Deny destructive operations, freeze writes during incident windows, and capture policy version and actor.
Runtime AI agent governance controls actions while they execute.
Runtime AI agent governance is the process of controlling autonomous agent actions while they are executing in production. Instead of relying only on prompts or pre-deployment tests, runtime governance intercepts tool calls — such as refunds, payments, data exports, database writes, or customer messages — and evaluates them against enforceable policy before the action reaches a real system.
HaltState provides this enforcement layer and preserves verifiable evidence of each decision, helping teams support incident response, internal controls, and regulatory alignment.
ALLOW
refund.create USD 126
APPROVAL_REQUIRED
refund.create USD 520
DENY
customer.pii.export
Built for evidence-driven AI governance.
Regulators and enterprise customers increasingly expect more than policy documents. They expect proof that controls operate at runtime. HaltState helps teams preserve action-level evidence for internal audits, incident response, and governance programs aligned to frameworks such as California SB 53 / TFAIA, the EU AI Act, New York RAISE, NIST AI RMF, ISO/IEC 42001, SOC 2 control narratives, and state-level automated decision-making laws.
California SB 53 / TFAIA
Transparency and incident-response evidence for runtime decisions.
EU AI Act
Human oversight, logging, and action-control support.
New York RAISE
Evidence for frontier AI safety and reporting direction.
NIST / ISO / SOC 2
Operational evidence and control narratives.
State ADM laws
Automated decisioning transparency support.
Framework-agnostic by design.
HaltState wraps tool calls and business actions at the SDK/API boundary, so teams can start with Python async functions and expand to agent frameworks, custom orchestration layers, and internal tools.
See a refund agent get stopped before execution.
The public Retail Agent Control Room shows a sanitized refund workflow: low-risk refunds pass, high-risk refunds require approval or get denied, and each decision produces a Proof Pack without exposing customer data.
See HaltState live
OpenClaw runtime safety
Put policy between OpenClaw and its tools.
Use pre-execution checks, approvals, scoped stops, and evidence logs before an OpenClaw agent can delete files, send messages, spend money, or change production systems.
Explore HaltState for OpenClawRespond first if an agent already caused damage.
Stop the runtime, disconnect risky integrations, preserve logs, rotate exposed secrets, and restore from systems of record before restarting automation.
Open the agent emergency checklist