> ## Documentation Index
> Fetch the complete documentation index at: https://docs.connex.hybridsphere.io/llms.txt
> Use this file to discover all available pages before exploring further.

# CONNEX AI Agent

> Autonomous reasoning engine, multimodal artifact generation, and sub-100ms precision query routing

The CONNEX AI Agent is an autonomous enterprise intelligence engine engineered for strategic reasoning, dynamic artifact generation, and multi-turn execution within strict token and security guardrails.

<CardGroup cols={2}>
  <Card title="Multimodal Deep Reasoning" icon="brain">
    Executes multi-step task planning, document synthesis, and domain logic resolution powered by flagship AI models.
  </Card>

  <Card title="Dynamic Artifact Sandboxes" icon="code">
    Generates interactive live components, executable code snippets, Recharts analytical dashboards, and Mermaid diagrams.
  </Card>

  <Card title="Agentic RCAG Grounding" icon="quote">
    In-text chunk citations (`[ref:filename#section]`) anchor all conclusions directly to verified enterprise source documents.
  </Card>

  <Card title="Sub-100ms Query Planner" icon="zap">
    1-turn 1-channel precision targeting minimizes latency and eliminates wasteful multi-turn reasoning loops.
  </Card>
</CardGroup>

***

## Agentic Reasoning Pipeline

The cognitive architecture follows a strict 4-stage pipeline ensuring deterministic business boundary enforcement and capital efficiency.

<Steps>
  <Step title="1. L1 Intent & Safety Gatekeeper (< 100ms)" icon="shield-alert">
    Evaluates prompt boundary, intercepts out-of-scope traffic, and validates intent without expensive model invocations.
  </Step>

  <Step title="2. Sub-LLM Query Planner" icon="git-branch">
    Targets exact sub-channels and retrieves authoritative context chunks from isolated VFS storage.
  </Step>

  <Step title="3. Flagship Deep Reasoner" icon="cpu">
    Executes grounded synthesis, structured data distillation, and citation fidelity verification.
  </Step>

  <Step title="4. Bidirectional Stream Typist" icon="sparkles">
    Emits real-time SSE token streams, dynamic artifact sandboxes, and interactive charts at 60 FPS.
  </Step>
</Steps>

***

## Technical Specifications

### 1. Token Capital Allocation Guardrails

All autonomous agent reasoning workflows operate within deterministic budgetary limits to eliminate infinite recursion and runaway compute costs.

| Metric / Guardrail | Value | Enforcement Layer | Purpose |
| :- | :- | :- | :- |
| **Max Turn Budget** | 5 Turns | Native Runtime Guard | Prevents runaway token bleed loops |
| **L1 Gatekeeper SLA** | \< 100ms | In-Memory / Flash LLM | Early escapes greetings and out-of-scope traffic |
| **Prompt Max Size** | 70,000 chars | WebSocket Frame Boundary | Prevents 1007 CloseFrame socket disconnects |
| **Citation Fidelity** | 100% Grounded | Post-Processor Filter | Eradicates speculative hallucinations |

### 2. WebSocket Streaming Protocol

Real-time bidirectional streaming delivers token updates directly to the client interface using structured event envelopes.

```typescript theme={null}
// Example: Subscribing to CONNEX live session stream
const socket = new WebSocket(`wss://connex.cloud/api/v1/agent/live?token=${sessionToken}`);

socket.onmessage = (event) => {
  const envelope = JSON.parse(event.data);
  switch (envelope.type) {
    case TEXT_DELTA:
      streamTypist.append(envelope.payload.delta);
      break;
    case ARTIFACT_EMIT:
      artifactRegistry.render(envelope.payload.artifact);
      break;
    case CITATIONS_RESOLVED:
      citationBadge.update(envelope.payload.sources);
      break;
  }
};
```


## Related topics

- [Frequently Asked Questions (FAQ)](/resources/faq.md)
- [AI Ecosystem 2.0 User Manual](/guide/workspace.md)
- [CONNEX | Docs Center](/index.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.