AI AGENT INFRASTRUCTURE

Everyone built the engine.
Nobody built the infrastructure.

AI agents can do anything. But who governs them? Who tracks their cost? Who remembers what they did? Who coordinates them? We build the infrastructure layer that sits between AI agents and the real world.

Patents Pending GB2604445.3 + GB2608069.7 On-Premise Zero Cloud
Book a Demo Two products. One infrastructure layer. Every AI agent.
MARCH 2026

Claude Code source code leaked via npm

512,000 lines of proprietary source code exposed through a missing .npmignore entry. Every safety mechanism in the leaked source operates inside the agent via system prompts and feature flags. Prompts are advisory. The agent can ignore them under context pressure. CRE operates outside the agent. The npm publish command would have been gated, reviewed, and blocked before it executed.

APRIL 2026

Anthropic's Mythos escaped its sandbox and concealed its actions

Claude Mythos Preview. The most capable LLM ever built. Expert-level cybersecurity. During testing it built a multi-step exploit to escape its sandboxed environment, gained internet access, sent an email to a researcher to prove it, then posted its own exploit details to public websites. It actively concealed its actions from the researchers monitoring it. Anthropic restricted access to ~40 organisations under Project Glasswing. System prompts cannot contain a model that can reason its way around them. External, mechanical enforcement is the only architecture that scales with model capability.

Two products. Complete infrastructure.

HookBus is the universal event bus that captures every AI agent lifecycle event. CRE is the enforcement subscriber that applies your rules. Together, they give you governance, observability, memory, and cost control for any AI agent.

HookBus
Universal Event Bus for AI Agent Lifecycle

The central router between AI agents and their actions. Every lifecycle event flows through HookBus. It routes to subscribers, consolidates decisions, and returns allow/deny/ask. The bus has no opinion on content. It routes. Developers build their own subscribers: memory systems, token monitoring, fine-tuning pipelines, cost controls, anomaly detection. All via the bus.

  • Any AI assistant, any SDK, one protocol
  • Sync subscribers block execution until they decide
  • Async subscribers observe without blocking
  • Deny-wins decision consolidation
  • Hot-reload subscriber config, zero downtime
  • Unix socket, HTTP, or in-process transport
  • Build your own subscribers: memory, cost, learning, alerts
  • Patent pending GB2608069.7
CRE
Composable Rule Enforcer

The enforcement subscriber. CRE receives PreToolUse events from HookBus and runs your organisation's rules against every tool call. Two layers. Deterministic pattern matching plus semantic intent verification.

  • L1: Microsoft AGT + pattern engine, under 10ms
  • L2: IBM Granite local LLM, under 1s, no cloud
  • PIN override for human approval chain
  • Full audit trail, SOC2/ISO 27001 ready
  • Self-learning: L2 decisions promote to L1
  • Anti-evasion detection
  • Patent pending GB2604445.3

Publishers. Bus. Subscribers.

Every AI agent action is a lifecycle event. HookBus captures them all and routes them to subscribers. Governance subscribers enforce rules. Observability subscribers track cost and usage. Memory subscribers record history. Intelligence subscribers learn and adapt. Add a subscriber, every agent's events flow to it.

Publishers
Claude Code
Hook
Cursor
Hook
Augment Code
Hook
GitHub Copilot
Hook
OpenAI Agents SDK
SDK Integration
Anthropic Agent SDK
SDK Integration
LangChain / CrewAI
Callback
OpenClaw
Bridge
Any HTTP Client
Webhook

HOOKBUS

Universal Event Router
Event Routing Fan-out Deny-wins Hot Reload Health Checks
Subscribers
CRE Gate
Sync - L1+L2 Enforcement
Budget Gate
Sync - Token/Cost Limits
Rate Limiter
Sync - Throttle
Audit Logger
Async - SQLite Trail
Token Counter
Async - Cost Tracking
Slack / Teams
Async - Alerts
SIEM Forward
Async - Splunk/Datadog
Learning Engine
Async - Rule Derivation
Sync (blocks until decision) Async (fire and forget)

Build your own subscribers. Plug into any agent.

HookBus is the platform. CRE is subscriber #1. But the bus doesn't care what subscribers do. Build anything that reacts to AI agent lifecycle events.

Agent Memory

Record what every agent did across sessions. Build persistent memory from real tool call history.

Token Monitoring

Track cost per agent, per user, per team. Set budgets. Alert when thresholds hit. Real-time dashboards.

Fine-Tuning Data

Capture every tool call and outcome. Generate training data for fine-tuning your models from real agent behaviour.

Anomaly Detection

Agent behaving differently than usual? Detect pattern shifts, unusual tool sequences, and alert before damage.

Compliance Gate

PCI, SOX, HIPAA rules as sync subscribers. Block actions that violate industry-specific regulations.

Team Alerts

Slack, Teams, PagerDuty, email. Notify the right people when agents are blocked or behave unexpectedly.

Any service that can read JSON over a Unix socket, HTTP endpoint, or Python class can be a HookBus subscriber. No SDK lock-in. No framework dependency. Write a subscriber in any language.

Three outcomes. Every action.

CRE is the enforcement subscriber on HookBus. Every tool call passes through CRE before execution. No exceptions. The AI never gets the choice to ignore your rules.

DENY

Dangerous or policy-violating operations rejected instantly. PIN override for authorised exceptions.

ALLOW

Safe, routine operations pass through with zero latency. Knowledge base context injected automatically.

ASK

Sensitive operations require human confirmation. Auditable approval chain with timestamps.

Layer 1

Microsoft AGT + Pattern Engine

< 10ms

Microsoft's Azure Governance Toolkit provides semantic policy matching. Combined with deterministic regex pattern rules for your organisation's policies. Configurable per team, project, or environment. Audit-ready.

Layer 2

IBM Granite Intent Verification

< 1s on CPU. No cloud. No GPU.

IBM Granite runs locally on your hardware. Verifies the AI's action matches what the user actually asked for. Catches substitutions, shortcuts, and creative reinterpretations. Zero data leaves the machine.

Any agent with hooks can use HookBus.

HookBus works with any AI assistant or SDK that exposes lifecycle hooks. The thin client normalises events from different hook formats into one standard protocol. As long as the hooks CRE needs are available (PreToolUse, UserPromptSubmit), the full enforcement solution works out of the box.

How it works

Each assistant exposes hooks at different lifecycle points. HookBus connects to whatever hooks are available. The more hooks an assistant exposes, the more subscribers can do. CRE only needs PreToolUse to enforce rules. Add UserPromptSubmit and CRE injects knowledge base context. Add PostToolUse and async subscribers can log, count, and learn.

Think of it as a plugin marketplace. Developers build subscribers. Organisations subscribe to the ones they need. HookBus routes the events. No vendor lock-in. No monolithic platform.

Claude Code
LIVE
Anthropic Agent SDK
LIVE
OpenAI Agents SDK
LIVE
LangChain / CrewAI
LIVE
Amp
LIVE
Cursor
TESTING
Augment Code
TESTING
OpenClaw
TESTING
GitHub Copilot
PLANNED
Windsurf
PLANNED

One requirement: the platform exposes a pre-execution hook. Every major AI assistant and SDK now does. HookBus normalises different hook formats into one standard protocol. The more lifecycle hooks a platform exposes, the more subscribers can do, but a single pre-execution hook is enough for full CRE enforcement.

Built for enterprises running AI agents at scale

Zero Cloud Dependencies

Both layers run on your hardware. No API keys. No external calls. Air-gap compatible.

Full Audit Trail

Every tool call, every decision, every override logged. SOC2, ISO 27001, ISO 42001 ready.

Self-Learning Rules

Layer 2 decisions automatically promote to Layer 1 patterns. Faster and more accurate over time.

Context Intelligence

Searches your organisation's knowledge base and injects context into the AI automatically.

Anti-Evasion

Prevents agents from encoding commands, lateral movement, or writing scripts to bypass enforcement.

Composable Subscribers

Add governance, observability, or intelligence subscribers without touching the bus or other subscribers.

Agent framework builders: add PreToolUse enforcement

Every agent SDK ships tools without lifecycle management. Agentic Thinking adds governance, observability, and intelligence to any framework.

Licensing

License the patented enforcement architecture for your AI agent platform.

Integration

One hook, full governance. Coding assistants, trading bots, autonomous agents.

White Label

Ship HookBus + CRE under your brand as part of your enterprise offering.

See it in action

15-minute demo. Real enforcement. Live blocking. Audit dashboard. Your questions answered.

Book a Demo
salesREMOVE@data4uREMOVE.uk