Guardrails & Safety
Continuous monitoring applied across all layers to ensure aligned and safe operation.
- Permissions
- Approval gates
- Content filtering
- Rate limits
- Human-in-the-loop (HITL) checkpoints
Perception Layer (Input)
Receives and structures inputs from the outside world before passing them to the cognitive core.
- User text / prompts
- API calls & Webhooks
- Sensor data
- System events & triggers
Reasoning Engine / LLM (Think)
The cognitive core determining if a goal is achieved, evaluating context, and deciding next steps.
- Chain-of-Thought processing
- ReAct methodology
- Plan-and-Execute evaluations
- Direct Answer generation
Memory System (Remember)
Short-term
- Conversation context
- Working state
Long-term
- Vector store (Embeddings)
- Episodic memory
- Learned patterns
Planning Module (Plan)
Triggered when complex reasoning is required. Breaks down macroscopic goals into executable sub-tasks.
Step 1 ➔ Step 2 ➔ Step 3
Tool Execution Layer (Act)
Executes the planned sub-tasks and returns results to the Reasoning Engine.
🔌 MCP
🌐 APIs
💻 Code Exec
🗄️ Database
📁 Files
🔗 External Services
Observability Layer (Observe)
Foundational monitoring layer capturing metrics, logs, and traces. Applies universally to ChatGPT, Claude, Copilot, Custom Agents, Multi-Agent Systems, and Autonomous Workflows.