Technology

Agentus Architecture

Deep dive into the technologies powering our autonomous agents. From cognitive engine to safety layer.

Agentus Cognitive Architecture

Multi-layered natural language processing and decision-making system

Cognitive Layer

Reasoning & Planning
Intent RecognitionChain-of-ThoughtDialogue State

Knowledge Layer

RAG & Memory
Vector SearchDeep ResearchContext Window

Action Layer

Tools & Execution
API IntegrationsFunction CallingResponse Generation

Safety Layer

Guardrails & Privacy
PII RedactionContent FilteringAccess Control

Cognitive Engine

The heart of the system. A three-step reasoning process allows agents to understand context and plan actions, not just respond with templates.

  • Perception: Analyze incoming messages and context.
  • Reasoning: Plan steps to achieve the goal (Chain of Thought).
  • Action: Execute tools or generate a response.
1
Input Analysis
User Intent Extraction
2
Reasoning Loop
Step-by-step Planning
3
Execution
Tool Call / Response

Dialogue State V2

Dialogue state stores the entire conversation context within the model window: user goals, filled slots, previous questions and answers. This allows the agent to continue conversations naturally.

  • Save slots (name, date, product) in context.
  • Context restoration after interruption.
  • Session lifecycle management and long-term memory.
1
Context Capture
Save User Intent & Slots
2
State Management
Track Conversation Flow
3
Context Restoration
Resume After Interruption

Slot Extractor Engine

Extracts structured data from natural language for API calls and tools. Works with fuzzy input.

  • Extract dates, times, locations, names, and amounts.
  • Data normalization (tomorrow at 5 -> 2025-11-25T17:00).
  • Validate required parameters before API calls.
1
NLP Extraction
Parse Natural Language
2
Data Normalization
Format & Validate
3
API Preparation
Structured Output

Deep Research

The agent gathers business knowledge automatically. Agentus can build a knowledge base by scanning your website, reviews, materials and forming structured product descriptions.

  • This mode helps you launch an agent quickly, even without a ready knowledge base.
  • You provide a website — we build a draft knowledge map for review and editing.
  • All data comes from public sources and your materials only.
1
Web Crawling
Scan Public Sources
2
Knowledge Graph
Build Structured KB
3
Review & Edit
Refine Knowledge Base

Telemetry Engine

Full transparency of agent operations. We log every decision so you can improve quality and debug scenarios.

  • Chain of Thought tracing.
  • Quality metrics and customer satisfaction (CSAT).
  • Alerts for anomalous behavior or errors.
1
Event Logging
Capture All Actions
2
Metrics Analysis
Track Quality & CSAT
3
Alerting
Detect Anomalies

Safety Layer

Enterprise-grade security. Your data is isolated and agent responses pass through safety filters.

  • GDPR & SOC2 Compliance-ready architecture.
  • Data isolation (Tenant Isolation) at the database level.
  • PII Redaction (automatic masking of personal data in logs).
1
Input Validation
Filter Malicious Content
2
Data Isolation
Tenant Separation
3
PII Redaction
Remove Sensitive Data

Try the technology in action

Create an agent for free and experience the quality of our cognitive engine.