Skip to content

Latest commit

 

History

History
45 lines (39 loc) · 2.09 KB

File metadata and controls

45 lines (39 loc) · 2.09 KB
name agent-platform
version 0.1.0
description Self-developing AI agent platform with multi-provider routing, RAG, skills, and workflow orchestration. Processes raw user inputs to autonomously build agent-based solutions.
author aleksandr
created 2025-02-01

Agent Platform — Master Specification

Platform Identity

  • Name: Agent Platform
  • Purpose: Autonomous agent development and execution platform
  • Philosophy: Bootstrap-first, cost-optimized, locally-first with cloud scale path

Core Capabilities

  1. Input Processing: Accepts markdown, CSV, DOCX, TXT, audio transcripts, conversation dumps
  2. Intent Understanding: Extracts user goals, constraints, and context from raw inputs via RAG
  3. Autonomous Planning: Generates implementation plans requiring human approval before execution
  4. Agent Development: Creates and deploys agent workflows, skills, and integrations
  5. Multi-Provider LLM Routing: Routes to cheapest viable model (local → cheap API → premium API)
  6. Skill Management: Discovers, loads, and executes AgentSkills (SKILL.md standard)
  7. Memory System: Short-term (context), long-term (semantic), episodic (history)
  8. RAG System: Local document ingestion with ChromaDB + Ollama embeddings
  9. MCP Integration: Connects to external tools via Model Context Protocol
  10. Workflow Orchestration: DAG-based multi-step workflow execution with checkpointing

Model Routing Priority

  1. Local Ollama (qwen2.5-coder:14b, deepseek-coder-v2:16b) — $0
  2. Gemini Flash / Deepseek API — budget tier
  3. Claude Sonnet / GPT-4o — quality tier (final outputs only)

Key Design Rules

  • All data stays local by default
  • Human approval gate before any execution phase
  • Every operation is logged with cost tracking
  • Skills follow Anthropic's SKILL.md standard (agentskills.io)
  • Workflows are defined as YAML, executed as DAGs
  • Memory consolidation runs after every 10 interactions

File Conventions

  • All Python files include type hints
  • All modules include docstrings
  • Configuration via .env (never hardcoded)
  • Tests mirror source structure under tests/