FIND-20260324-018 — 724-office — Self-evolving AI Agent System (Pure Python, 26 tools, MCP/Skill plugins)

724-office is a production-ready self-evolving AI agent system written in ~3,500 lines of pure Python with zero framework dependencies. It features 26 built-in tools across 8 files, a three-layer memory architecture (session + LLM-compressed long-term + vector retrieval via LanceDB), MCP/Skill plugin system (JSON-RPC over stdio/HTTP), OpenAI-compatible function calling, persistent cron scheduling, daily self-repair diagnostics, multimodal support (image, video, voice, files), and runtime tool creation (agents can write and load new Python tools dynamically). Reaches 844 stars within 7 days of creation (March 17-24, 2026). Edge-deployable on Jetson Orin Nano. Shared via @ihtesham2005 on X (tweet ID 2036186258918220290). Direct architectural analogue to the ODS ADLC agent system: single-file tool registry, skill plugin model, persistent memory, self-repair, MCP protocol integration.

Directly relevant to the ODS ADLC pipeline architecture. The three-layer memory model (session + compressed + vector) could enhance agent continuity in the ADLC supervisor. The MCP/Skill plugin pattern mirrors the ODS ~/.claude/skills/ system and could inform standardization of skill interfaces. The runtime tool creation capability (agents writing their own tools) is an advanced autonomy pattern worth tracking for ADLC v3+. The self-repair and daily diagnostics loop could inform the ODS agent health-check system. The zero-dependency pure Python approach is a useful reference for the ODS ops scripts. Most immediately actionable: the three-layer memory architecture and the MCP plugin pattern for write-finding.sh / write-review.sh equivalents.

724-office — Self-evolving AI Agent System (Pure Python, 26 tools, MCP/Skill plugins)

Source

ODS Impact

Security Review

Tags