Question 1

What is Agent Runtime Governance (ARG)?

Accepted Answer

Agent Runtime Governance is the discipline of governing what an AI agent does at runtime, before and during execution. It authorizes every action, delegation, and tool invocation against declared intent, delegated authority, and current policy, enforces the decision at the moment of execution, contains a running agent when it goes wrong, and records what happened as signed execution lineage. ARG sits between identity infrastructure (IAM) and the agent frameworks that execute work.

Question 2

How is ARG different from IAM?

Accepted Answer

IAM answers "who is this entity and what can they access?", evaluated at authentication time with static roles and periodic reviews. ARG answers "should this agent take this action right now, under whose delegated authority, and does it align with an approved, time-boxed business objective?", evaluated on every action, in real time, with full delegation and intent context. IAM governs access. ARG governs execution.

Question 3

Is ARG just ABAC with extra steps?

Accepted Answer

No. ABAC evaluates attributes at request time: subject, resource, action, environment. ARG evaluates runtime context that ABAC has no model for: declared intent, active goals with time limits and action limits, delegation chains with strict-subset attenuation, plan-act-observe lifecycle state, and drift between declared intent and actual behavior. It also acts on a running agent, which a policy evaluation pattern cannot do. ABAC is a policy evaluation pattern. ARG is an architectural layer.

Question 4

Does ARG replace IAM or PAM?

Accepted Answer

No. ARG complements them. IAM continues to manage identity lifecycle and access provisioning. PAM continues to manage privileged access workflows. Agent Runtime Governance adds the layer that neither provides: intent validation, delegation with strict-subset attenuation, continuous policy enforcement, real-time containment, and signed execution lineage for autonomous agents.

Question 5

What is the minimum to get started with ARG?

Accepted Answer

Start with agent discovery and registry: know what agents and MCP servers exist, who owns them, and assign each a verifiable identity. Then add intent declaration and runtime policy enforcement. These three capabilities move you from Level 0 (Unmanaged) toward Level 2 (Controlled) on the five-level Agent Runtime Governance maturity model. Our position paper details all seven minimum requirements.

Question 6

What does Watchlight AI do?

Accepted Answer

Watchlight AI introduced and defined Agent Runtime Governance as a distinct architectural discipline in February 2026 and builds Watchlight AI Beacon, the enterprise runtime control plane for AI agents. Beacon decides whether an agent may act, enforces that decision before the action runs, and proves what happened. It is available now, ready to install in your environment, on-premises or air-gapped. Practitioner-led services (executive briefings, workshops, and implementation) help teams put Agent Runtime Governance into practice.

Question 7

Who is Watchlight AI for?

Accepted Answer

We work with CISOs, CTOs, enterprise architects, and security leaders at organizations deploying or planning to deploy AI agents in production. If your agents access enterprise systems, make autonomous decisions, or orchestrate multi-step workflows, you need Agent Runtime Governance. Watchlight AI Beacon provides it, and our services help you design for it.

Question 8

What industries do you work with?

Accepted Answer

We focus on industries where autonomous agents are moving into production, including financial services, healthcare, technology, government, and professional services. Agent Runtime Governance is industry-agnostic: any organization deploying autonomous agents needs governance infrastructure, regardless of sector.

Question 9

How is Watchlight AI different from IAM vendors or AI governance platforms?

Accepted Answer

IAM vendors focus on identity lifecycle and access provisioning: they govern who can access what. Broad AI governance platforms focus on model inventories, bias detection, and compliance checklists. Watchlight AI governs the authority behind every agent action at runtime, before and during execution. Beacon covers the architectural layer between identity infrastructure and agent frameworks that neither category covers: deterministic per-action authorization, delegated authority with strict-subset attenuation, real-time enforcement effects, drift detection that quarantines, and signed execution lineage, enforced in-process and on the wire.

Question 10

How can we engage with Watchlight AI?

Accepted Answer

Two ways. The product: Watchlight AI Beacon, the enterprise runtime control plane for AI agents, available now and ready to install in your environment, on-premises or air-gapped. And three practitioner-led services, calibrated to where you are in your agent program: a half-day executive briefing, hands-on workshops and architecture engagements, and an implementation engagement that takes a blueprint into production.

Question 11

How does an engagement typically start?

Accepted Answer

Most engagements begin with a governance readiness assessment: a structured evaluation of your current agent deployments, governance maturity, and security posture. We identify ungoverned agents, map delegation chains, assess policy gaps, and deliver a prioritized roadmap. From there, engagements can progress into a Beacon deployment, architecture, implementation, or ongoing advisory depending on your needs.

Question 12

Do you offer assessments or maturity evaluations?

Accepted Answer

Yes. Our governance maturity assessment maps your organization against a five-level model: Level 0 (Unmanaged), Level 1 (Identified), Level 2 (Controlled), Level 3 (Governed), and Level 4 (Optimized). It identifies which of the 12 Principles you have addressed, where gaps exist, and what to prioritize next. Most organizations discover they are at Level 0 or Level 1. A free 12-question self-assessment is available on our site.

Question 13

Can you work with our existing IAM, PAM, and SIEM stack?

Accepted Answer

Yes. Agent Runtime Governance is designed to complement your existing security infrastructure, not replace it. We integrate with your IAM (Okta, Entra ID, Ping, etc.), PAM (CyberArk, Delinea, etc.), SIEM (Splunk, Sentinel, etc.), and policy engines. ARG adds the agent-specific runtime layer on top of what you already have.

Question 14

Do you build custom solutions or provide advisory only?

Accepted Answer

Both. The product, Watchlight AI Beacon, is the runtime control plane. Around it we provide advisory and architecture services (assessment, strategy, and governance design) and hands-on implementation that deploys Beacon into your environment. We tailor the engagement to your team and maturity level.

Question 15

What agent frameworks do you support?

Accepted Answer

Watchlight AI Beacon governs agents across the frameworks enterprises actually run, through drop-in plugins: LangGraph, the Claude Agent SDK, Claude Code, Google ADK, AWS Bedrock Agents, the Microsoft Agent Framework, Pydantic AI, DeepAgents, and OpenClaw, plus MCP servers and custom or in-house agents. Coverage per framework varies by what its lifecycle exposes. The on-the-wire proxy is framework-agnostic: it governs every outbound request regardless of how the agent produced it, so enforcement is not coupled to any one orchestration tool. For agents running in NVIDIA OpenShell sandboxes, Beacon also integrates with OpenShell directly, deciding every request inside OpenShell’s supervisor.

Question 16

Does ARG require changes to our existing agent code?

Accepted Answer

Often not. Beacon enforces at two layers that act on one policy decision. An on-the-wire proxy governs every outbound request without agent code changes: agents make standard requests through it, and policy is evaluated on the wire. In-process plugins add semantic, plan-stage governance for supported frameworks with a drop-in install (a few configuration values, no rewrite). Deeper intent declaration and goal registration can be adopted incrementally.

Question 17

What decisions can Beacon return, and what happens when an agent goes wrong mid-run?

Accepted Answer

Every action receives one of three verdicts: Allow, Require approval, or Deny. Beyond the per-action verdict, Beacon can act on a running agent with real-time enforcement effects: stop the run in progress, quarantine the agent, sever the delegation subtree beneath it, or revoke its authority across the fleet. When an agent drifts from its declared plan, it is quarantined automatically, at machine speed, rather than queued as an alert for a human to triage later.

Question 18

Is a language model involved in the authorization decision?

Accepted Answer

No. Every decision is evaluated against formal, versioned policy by a deterministic engine. No language model sits in or feeds the decision, so a persuasive justification cannot talk its way past a rule, and the same policy yields the same verdict every time. That is what makes the control auditable.

Question 19

How does runtime policy enforcement work without adding latency?

Accepted Answer

Runtime policy evaluation is designed for production workloads and runs inline at the point of action. Policies are expressed in formal policy languages (not natural language) and evaluated by a purpose-built engine that selects the relevant policy subset for each request rather than scanning every policy. The overhead is designed to be comparable to an API gateway check.

Question 20

What is intent-based authorization?

Accepted Answer

Intent-based authorization shifts the question from "what can this agent access?" to "why is this agent taking this action?" Every agent declares its purpose (why it exists), goal (what it is trying to achieve, with scope and time limits), and intent (what it is doing right now). Authorization decisions evaluate all three against the requested action. An agent declaring "research" intent cannot perform a delete operation, because the intent and action are inconsistent.

Question 21

What are delegation chains?

Accepted Answer

A delegation chain is the trust path from the human who authorized an action through every agent in the execution path. When a user asks an orchestrator agent to generate a report, and that agent delegates to a data retrieval agent, the chain traces: Human → Orchestrator → Data Agent. At each link the child receives a strict subset of the parent's authority, and any attempt to widen it is denied before the action runs. The whole chain is recorded as signed, tamper-evident execution lineage and is auditable end to end.

Question 22

How does ARG relate to NIST AI RMF, ISO 42001, and the EU AI Act?

Accepted Answer

ARG is complementary to these standards. NIST AI RMF provides organizational risk management. ISO 42001 provides management system requirements. The EU AI Act provides regulatory requirements. ARG provides the agent-specific prescriptive depth they lack: the concrete architectural patterns for identity, authorization, containment, execution lineage, secrets management, and multi-agent coordination that broader standards do not prescribe. Use them together: standards for compliance frameworks, ARG for implementation.

Question 23

Can ARG help us pass an AI-related audit?

Accepted Answer

Yes. Beacon produces the evidence auditors need as signed, tamper-evident execution lineage: which agent took which action, under what business objective, authorized by whom, through which delegation chain, evaluated against which policy version, and what the outcome was. This is the record that NIST AI RMF, ISO 42001, SOC 2, and EU AI Act record-keeping reviews increasingly ask for, and that most organizations cannot provide today because they lack runtime governance infrastructure for their agents.

Question 24

How does ARG support SOC 2 or regulatory compliance?

Accepted Answer

SOC 2 requires demonstrable controls over system access, change management, and monitoring. ARG provides these controls for AI agents: verifiable identity and registry (access controls), runtime policy enforcement (change management), and signed execution lineage with drift detection (monitoring). For regulated industries, ARG also supports data governance requirements by ensuring agents only access data consistent with their declared purpose and approved goals.

Question 25

How do we know if we need Agent Runtime Governance?

Accepted Answer

If you are deploying AI agents that access enterprise systems, make autonomous decisions, delegate to other agents, or invoke tools dynamically, you need ARG. Ask yourself: Can you enumerate every agent and MCP server in your environment? Can you explain why any agent was allowed to take any action? Can you trace an agent's authority back to a human decision, and prove a sub-agent never exceeded its parent? Can you stop a running agent in real time? If you cannot answer yes to all four, you have governance gaps that ARG addresses.

Question 26

What does a governance maturity assessment look like?

Accepted Answer

We conduct a structured evaluation. We inventory your AI agents, map their access patterns and delegation chains, assess your current governance controls against the 12 Principles, score your maturity level (0 to 4), and deliver a prioritized roadmap with specific, actionable recommendations. The output is an executive-ready report and a detailed technical plan.

Question 27

Where can we read more?

Accepted Answer

Start with our pillar page on Agent Runtime Governance for the category definition. Then read our position paper "Why Agent Runtime Governance Is the Missing Layer in IAM" for the full argument. For the complete technical framework, download "12 Non-Negotiable Principles for Agent Runtime Governance." Our blog covers individual principles and topics in depth.

Frequently Asked Questions

Agent Runtime Governance

Watchlight AI

Services & Engagement

Technical & Architecture

Compliance & Standards

Getting Started

Ready to Govern Your AI Agents?

We value your privacy