a framework for building agents that work.

skillgraph is an experimental agentic framework built to solve the problems i kept running into with existing tools. it's open source (apache 2.0), it's cheaper to run, and it gives you actual control over what your agent does.

instead of giving agents a mess of low-level tool functions to fumble through, skillgraph uses skills - more sophisticated units that handle specific tasks with their own logic and workflows. they can orchestrate multiple tools, manage state, and even act as subagents when needed.

why i built this

i was frustrated with existing frameworks. every tool i tried had the same problems:

  • •cost: running agents with traditional tool-calling is expensive. every tool call wastes tokens on planning, coordination, and back-and-forth.
  • •control: hard to control what the agent does when you're just giving it functions and hoping it picks the right ones.
  • •complexity: multi-turn workflows are painful to implement. want to book a ticket with confirmations and payment? good luck.
  • •tool-calling sucks: agents spend forever deciding which tools to call and in what order, burning tokens the whole time.

so i built skillgraph. it's not production-ready yet, but it works, and the core ideas are solid. use it, break it, tell me what's wrong.

what makes it different

skills handle their own logic

instead of the agent burning tokens to figure out which tools to call and how to chain them, skills just do their job. the agent delegates, the skill executes. less overhead, less waste.

native multi-turn support

need a workflow that spans multiple messages? skills can enter 'skill mode' and take control of the conversation until the task is done. no awkward workarounds.

subject-object memory architecture

replaces complex rag systems with simple state tracking. the subject tracks what the user wants (goals, constraints, preferences). the object tracks what's being discussed right now. fast, cheap, and it actually works.

smart caching that saves money

anthropic prompt caching for static content (89% cost reduction on system prompts) + redis for conversation history (50ms → 5ms retrieval). multi-layer optimization that compounds.

built-in reliability

llm fallback chains (beta → alpha → retry) mean 99.9% uptime even during partial outages. graceful degradation everywhere.

open source. apache 2.0.

use it commercially, modify it, do whatever you want. the code is on github, the license is permissive, and contributions are welcome.