YAML-driven agent orchestration for the LLM-agnostic

What it does

AgentForge is a low-code framework for building autonomous AI agents and multi-agent systems. You define agents, their personas, and memory-backed workflows in YAML files called Cogs, then run them against whichever LLM fits the task—OpenAI, Anthropic, Google, or local models via Ollama and LMStudio.

The interesting bit

The Cog abstraction is the real bet: it treats multi-agent orchestration, branching logic, and memory as declarative infrastructure rather than imperative code. Memory nodes get declared in Cogs and injected into agent prompts automatically, which is the kind of glue that usually ends up as spaghetti in hand-rolled systems.

Key highlights

• Declarative Cogs: YAML-configured workflows that compose agents with branching logic and memory
• Model-agnostic by design: Different agents can run different LLMs in the same system
• Integrated memory: ChromaDB-backed vector storage, declared in Cogs and auto-available to agents
• Live prompt editing: Modify prompts without restarting the runtime
• Persona system: Separate agent identity, style, and reusable knowledge into their own config files

Caveats

• The existing Actions and Tools system is deprecated and will be replaced with an MCP-based implementation; current tool integrations are in limbo
• The project is actively seeking a volunteer UI/UX contributor, which suggests the frontend story is currently thin or nonexistent

Verdict

Worth a look if you’re building multi-agent prototypes and want to experiment with topology without drowning in orchestration code. Skip it if you need mature tool integrations today or a polished interface—this is backend plumbing with ambition.