🧠 AgentOS — System Architecture
The AI-Native
Operating System
AgentOS is not a group of agents — it is a stack of eight layers that enables agents to collaborate, coordinate, and continuously improve over time.
AgentOS = A multi-layer AI operating system that coordinates interfaces, workflows, and specialized agents to autonomously execute and improve tasks over time. It is not a tool — it is an AI-native operating environment.
8
System Layers
15+
Specialized Tools
6
Agent Roles
🔥 Key Insight
AgentOS is an operating environment,
not a collection of tools.
not a collection of tools.
What you've built is a system that has human interfaces, task coordination, multi-agent intelligence, execution pipelines, shared memory, a personal knowledge compiler, and a growth layer — working together as one.
01
Full System Architecture
8-layer flow
👤 User
Human input
↓
🌐 Interface Layer
Slack · OpenClaw · NanoClaw · Pi · Droid · Crush
↓
📋 Coordination Layer
Claw-Kanban · VibeKanban · Task state management
↓
🧭 Planner / Orchestrator
Manus · Node.js Orchestrator
↓
🤖 Intelligence Layer (Agents by Role)
Perplexity · Vane · Gemini CLI · Claude Code · Codex · Aider · Ollama/Qwen
↓
⚙️ Execution Layer
Aider · Codex · CLI tools · Local scripts
↓
🧠 Memory Layer (Shared Brain)
/data filesystem · Qdrant vector DB · Postgres state · Log files
↓
🧠 Cortex (Personal Knowledge Base)
raw/ → LLM Compiler → cortex/ .md wiki · Obsidian frontend · Linting
↕
🌐 Culture Layer
Claw Empire · Crush · Identity · Virality · Narrative
02
The Eight Layers — Detailed
Each layer's role and tools
Layer 1 · Interface
Human ↔ System
Entry point for all users. Handles conversational interaction, intent capture, and human feedback loops. These are not agents — they are interaction surfaces.
Slack
OpenClaw
NanoClaw
Pi
Droid
Crush
Layer 2 · Coordination
Task Systems
Tracks tasks, manages workflow visibility, and maintains project state. Shows what agents are doing. Enables async collaboration between human and AI.
Claw-Kanban
VibeKanban
Todo/Doing/Done
Layer 3 · Planner
Orchestrator
Breaks goals into tasks, routes to agents, controls loops and retries. Manus acts as a hybrid planner + executor — the decision-making core of the system.
Manus
Node.js Orchestrator
Task Router
Layer 4 · Intelligence
Core Agents (by Role)
Specialized AI agents each assigned a specific role: Cloud Researcher, Private Researcher, Builder/Executor, Critic/Reviewer, or Local Fast Intelligence. No single agent does everything.
Perplexity
Vane
Claude Code
Gemini CLI
Codex
Aider
Ollama/Qwen
Layer 5 · Execution
Tooling Environment
Actually runs the code, modifies files, and executes workflows. The execution layer receives instructions from the Intelligence Layer and performs the real-world actions.
Aider
Codex
Local scripts
CLI tools
Layer 6 · Memory
Shared Brain
Maintains shared context across all agents. Provides long-term knowledge storage and retrieval for RAG. Vane connects directly to Qdrant for proprietary dataset search alongside web results. All agents read and write to the same memory layer.
/data filesystem
Qdrant vector DB
Postgres state
Log files
Vane Embeddings
Layer 7 · Cortex
Personal Knowledge Base
LLM-compiled knowledge wiki inspired by Karpathy's architecture. Raw research from Perplexity and Vane flows into raw/, gets compiled into interlinked .md articles, and is periodically linted for consistency. Replaces RAG for personal-scale knowledge (100–10K docs). Viewable in Obsidian.
raw/ ingest
LLM Compiler
cortex/ wiki
Obsidian
Linting agent
Layer 8 · Culture
Growth Layer (NEW)
This is what turns AgentOS from a tool into a movement. Drives identity, engagement, virality, and narrative. Makes the system feel alive — not just functional.
Claw Empire
Crush
Identity
Virality
Narrative
03
Intelligence Layer — Agent Roles
Researchers · Builders · Critics · Local
🌐 Cloud Researcher
External knowledge, market research, and real-time web data. Queries the live internet for ground truth. Used when data is public and current.
Perplexity Computer
Gemini CLI
🔒 Private Researcher
Analyzes sensitive business data — financials, agreements, contracts, proprietary datasets. Runs fully local via SearxNG + local LLMs. No data leaves your infrastructure.
Vane
SearxNG
Ollama/Qwen
🛠️ Builders / Executors
Code generation, implementation, and system building. Aider = strong interactive coding. Qwen = cheap, fast local builds.
Claude Code
Codex
Gemini CLI
Aider
Ollama/Qwen
🧪 Critics / Reviewers
Evaluate outputs, provide feedback, and score quality. Gate keeper — only passes work scoring 8+ out of 10.
Claude
Gemini CLI
⚡ Local Fast Intelligence
Drafts, summaries, preprocessing, cheap iterations. Acts as an intelligence cache — reduces cost for repetitive tasks.
Ollama Web UI
Qwen 3B–7B
04
Execution Flow — End to End
11-step orchestration
1
User → Slack / Claw UI Interface Layer
User submits a goal, request, or task via the interface surface of their choice. Intent is captured and queued.
2
Task Created → Claw-Kanban Coordination Layer
Task enters the kanban board with status "todo". Workflow visibility enabled — both human and agents can see state.
3
Manus → Plan + Decompose Planner
Manus breaks the goal into atomic subtasks and routes each to the appropriate agent. Sets priority and dependencies.
4a
Perplexity → Cloud Research Intelligence · Cloud Researcher
Perplexity gathers real-time external information, market data, and factual grounding from the public web.
4b
Vane → Private Research Intelligence · Private Researcher
Vane analyzes sensitive business data — financials, contracts, agreements — using local LLMs and SearxNG. No data leaves the infrastructure. Custom actions can call your ML pipeline sidecar.
5
Qwen → Summarize Intelligence · Local Fast
Ollama/Qwen compresses research output into a structured brief. Cheap, local, fast — reduces cost before expensive models act.
6
Builder (Aider / Codex) → Implement Execution Layer
Aider or Codex implements the solution — writes code, modifies files, runs scripts. Output is concrete and testable.
7
Claude → Critique Intelligence · Critic
Claude reviews the implementation against the original goal. Scores output 1–10. Returns structured feedback with improvement notes.
8
Manus → Decide: Iterate / Escalate / Complete Planner
Score < 8: loop back to builder with critique. Score ≥ 8: mark complete. Escalate to human if blocked after 3 attempts.
9
Results → Memory Layer Memory
Output, critique scores, and context stored in Qdrant + Postgres. Available for all future agents to reference via RAG.
10
Cortex → Compile to Wiki Cortex
Research output and results flow into raw/. LLM compiler distills into structured .md wiki articles with backlinks and summaries. Knowledge compounds — every query adds to the base.
11
Output Displayed + Shared Interface + Culture
Results surface via Slack/Claw UI. Culture layer amplifies successful outputs — sharing, identity, and narrative.
05
Role Clarification Matrix
What each layer is and does
| Layer | What It Is | Examples | Key Function |
|---|---|---|---|
| Interface | User interaction surfaces | Slack, Pi, Droid | Entry points — where humans talk to the system |
| Coordination | Task tracking systems | Claw-Kanban | Makes work visible; manages state for projects |
| Planner | Decision maker & router | Manus | Decomposes goals; routes to right agents; controls loops |
| Intelligence | Thinking agents by role | Claude, Gemini, Perplexity, Vane | Specialized reasoning — cloud research, private research, build, critique, summarize |
| Execution | Doing the actual work | Aider, Codex | Writes code, modifies files, runs scripts |
| Memory | Shared brain / context store | Qdrant, Postgres, /data | Long-term storage; RAG retrieval; cross-agent context |
| Cortex | LLM-compiled knowledge base | raw/ → wiki/ → Obsidian | Compiles research into persistent, queryable .md wiki — replaces RAG at personal scale |
| Culture | Growth & identity layer | Claw Empire, Crush | Turns the system into a movement — virality, narrative |
06
What You've Built
System capability checklist
✓
Human interfaces — Multiple entry points via Slack, Claw UIs, and conversational agents
✓
Task coordination — Kanban-based workflow visibility and async human-AI collaboration
✓
Multi-agent intelligence — Specialized agents: cloud researcher, private researcher, builder, critic, local fast
✓
Execution pipelines — Aider and Codex implement outputs from intelligence layer
✓
Shared memory — Qdrant vector DB + Postgres for persistent cross-agent context
✓
Growth / viral layer — Culture layer builds identity and narrative around the system
✓
Quality gates — Critique loop scores work 1–10; only score ≥ 8 advances
✓
Cortex knowledge base — LLM-compiled wiki turns disposable research into persistent, compounding intelligence
✓
Self-improving loops — Failed tasks revise and retry; successes stored in memory for future reference
07
Explore Related Docs
Dig deeper into each layer