Inside the AgentProtect demo
We have published a customer-facing AgentProtect demo that shows, in one screen, how events from AI agents, chat assistants, coding assistants, and tool-using runtimes move through policy, approval, knowledge, and audit subscribers before becoming runtime evidence.
The point of the demo is to show the event trail around any AI runtime: what the assistant, agent, coding tool, SDK, CLI, or automated workflow tried to do, which subscriber saw the event, whether the action was allowed, denied, or paused for approval, and what evidence remains afterwards.
That is the enterprise problem. Regulated teams do not just need a safer prompt. They need a runtime layer that can prove what happened.
What the demo is showing
The demo is a simplified walkthrough of the same architecture used by the AgentProtect suite. An AI runtime emits lifecycle events. That runtime might be an autonomous agent, a chat assistant, a coding assistant, a CLI, an SDK, or a workflow runner. HookBus receives those events. Enterprise subscribers inspect, enrich, approve, deny, redact, remember, or audit the action.
In the demo, you choose a scenario, run it, then click the highlighted event. The highlighted event is the important part: it shows the decision point and the evidence envelope that a governance, audit, or workflow system would need.
The scenarios are deliberately concrete:
- A safe action that is allowed and recorded.
- A dangerous tool call that is blocked before execution.
- An external email that is paused for human approval.
- A reasoning example where exposed model reasoning metadata is retained when the provider makes it available.
How this maps to the actual suite
AgentProtect is not one giant policy engine. It is a subscriber bundle. That matters because enterprises already have different systems for identity, workflow, logging, knowledge, and audit. HookBus sits at the event layer and lets those controls subscribe to the same agent runtime evidence.
AgentRegistry
Correlates API-key use, user and device approval, runtime registration, and evidence production so sanctioned AI does not become unmanaged AI.
Composable Rule Enforcer
Evaluates the agent action before execution using deterministic controls first, with optional semantic review for borderline cases.
AgentFlow Approval
Pauses ask-level actions, sends the approval request to the right process owner, and records the approve or deny decision as runtime evidence.
AgentAuditor
Records chained, exportable runtime evidence for internal audit, assurance, and incident review.
Why the event log matters
The demo is built around the event log because the event log is where enterprise governance becomes inspectable. If an agent attempts a tool call, the useful evidence is not just the final answer. It is the chain of facts around the action:
- What prompt or session caused the action?
- What tool or model call was about to happen?
- Which subscriber inspected it?
- Was the decision allow, deny, ask, redact, enrich, or audit?
- If a human approved it, who or what role approved it?
- If the action was delegated, which agent was handed the decision?
- What metadata, reasoning summary, redaction, or audit record remains?
That is what boards, CISOs, DPOs, internal audit, insurers, and regulators will need to understand. Not a marketing claim that the AI system was safe, but a runtime record showing what it did and which controls acted.
The demo is simulated. The architecture is the product. The public page does not process customer prompts and does not send telemetry. It is a guided example of how the real AgentProtect subscriber suite routes events inside an on-premise, VPC, or air-gapped deployment.
Where AgentHook fits
AgentHook defines the open runtime evidence vocabulary. HookBus is Agentic Thinking's event bus and enterprise implementation path. The customer demo is on agenticthinking.uk/demo because customers need to see the commercial workflow: enterprise subscribers, approval paths, audit, and deployment support.
The standards work remains separate. AgentHook is the neutral specification. AgentProtect is the product suite we deploy and support for regulated teams.
What to do next
If you are evaluating AI agents, chat assistants, coding assistants, or tool-using AI workflows in a regulated environment, use the demo as a conversation starter. Ask where your current AI stack records policy decisions, human approvals, model metadata, and audit evidence today.
If the answer is scattered across prompts, vendor logs, screenshots, and Slack threads, you do not have runtime governance yet. You have policy documents around an ungoverned runtime.
AgentProtect generates runtime evidence that can support governance, audit, and compliance programmes. It does not certify legal or regulatory compliance on its own. Compliance determinations remain the responsibility of the implementing organisation and its advisers.
← Back to blog