Arkos

A cognitive AI operating system. Local. Modular. Vision-first.

Arkos is not another LLM wrapper. It's an orchestration engine that coordinates specialized local models to think, plan, simulate, execute, and learn — autonomously.

Core Idea

Standard LLMs predict the next token. They don't know what they're building.

Arkos flips this:

Prompt → Vision → Simulate Users → Extract Goals → Plan → Execute → Evaluate → Learn

The intelligence lives in the architecture, not the model size.

Architecture

User Prompt
     │
     ▼
Vision Generator        ← forms a design before any code is written
     │
     ▼
Scenario Simulator      ← imagines real users interacting with the product
     │
     ▼
Goal Extractor          ← derives goals from simulated friction + delight
     │
     ▼
Feasibility Checker
     │
     ▼
Planner                 ← builds ordered task graph
     │
     ▼
Execution Pool          ← parallel specialized workers (code, debug, docs)
     │
     ▼
Evaluator / Critic      ← scores output; triggers adaptive context escalation
     │
     ▼
Memory                  ← persistent lessons + patterns

Key Principles

• Vision before action — no worker runs without a plan
• Adaptive context — minimal context by default, escalate on low confidence
• Bounded retries — max 3 attempts, then escalate (no infinite loops)
• Parallel execution — independent tasks always run simultaneously
• Local first — no cloud, no API keys, no data leaving the machine
• Memory mandatory — every run produces artifacts the system learns from

Tech Stack

• Orchestrator: Rust / Node.js
• Models: Ollama (Mistral, Qwen-Coder, CodeLlama, Phi)
• Memory: Qdrant vector DB
• Tools: file read/write, shell exec, semantic codebase search

Status

🚧 Early development — see BIBLE.md for full architecture spec.

Docs

• BIBLE.md — full architecture specification
• docs/ — module-level docs (in progress)