One agent stack, three doors: SDK, desktop, and an enterprise gateway

What it does

AgenticX is a multi-agent platform built around a Python core: an SDK, an agx CLI, and a FastAPI Studio server. On top of that sit two product forms — Near Desktop, a local-first Electron + React multi-agent workspace with group chat, avatars, and an embedded terminal, and an Enterprise tier with a web portal and a Go AI Gateway handling auth, policy, quotas, and audit. The core covers the usual agent-platform checklist: 15+ LLM providers, MCP Hub, hierarchical memory, GraphRAG knowledge bases, a skill system, and Docker/subprocess sandboxes.

The interesting bit

The README is unusually honest about its own boundaries: it states plainly that the Enterprise Gateway is a compliance relay, not an Agent Runtime, that Cluster/HA is still planned, and that some storage adapters are placeholders. There’s also a Meta-Agent “CEO dispatcher” that routes group chats and spawns sub-agents with concurrency limits — plus skill self-evolution, where runtime tool-call observations get distilled into new skills by a background LLM review session.

Key highlights

• Embeddable ReActAgent primitive with typed event streaming, parallel tool execution, and zero coupling to the desktop app
• Unified offloading keeps large tool results out of live context via filesystem-backed references
• Skill lifecycle includes a dangerous-pattern security scan gate and fuzzy patching
• Long-horizon coding loop with a disk-backed project state machine and stall self-healing
• A2A and MCP protocol support, plus Feishu/WeChat sidecars

Caveats

• The Studio path doesn’t yet route every operation through the full SafetyLayer — hooks and path protection carry the load for now
• Enterprise Edge Agent is a non-default MVP; Cluster Runtime hasn’t started
• Scope is enormous — some third-party storage adapters are unwired placeholders

Verdict

Worth a look if you want a batteries-included, local-first agent workspace with a real SDK underneath and can tolerate a young, sprawling codebase. If you wanted a thin, minimal agent loop, this is the opposite of that.