skillgraph
a framework for building AI agents that work
15 followers
a framework for building AI agents that work
15 followers
An experimental agentic framework that replaces traditional tool-calling with skills. Built for developers who want more control, teams and enterprises who need agents that are context grounded, lower costs, and better workflows. Open source under Apache 2.0.
Hey Product Hunt,
I'm Tejas, the founder of Fanpit, a platform that helps turn ideas to brands and scale them and this puts us at the heart of consumer AI. To do that right, we need AI agents that actually understand context and can handle complex workflows without burning through our budget.
Existing frameworks weren't cutting it. Rigid tool-calling patterns, expensive token usage, no real context awareness. So we built skillgraph to solve our own problems.
The core idea is skills instead of tools. Traditional frameworks give agents a mess of low-level tools to fumble through. skillgraph uses skills - autonomous units that handle entire workflows. The agent delegates, the skill executes. Less overhead, less waste.
We also built subject-object state tracking. Instead of complex RAG systems, we track the subject (user goals, constraints, preferences) and the object (current topic being discussed). Fast utility LLM does the analysis. 50ms retrieval from Redis. It works.
The results: 50-67% fewer LLM calls in our testing, 89% cost reduction on system prompts through prompt caching, multi-turn workflows that actually work, and 99.9% uptime with LLM fallback chains.
We've been building on this and we're about to take it to production at Fanpit.
Then we realized this isn't just useful for customer experience. Healthcare teams managing patient workflows. Finance teams processing documents. Support teams handling complex queries. Anywhere you need agents that understand context without destroying your budget.
So we're open sourcing it. Apache 2.0. No strings.
Website: https://skillgraph.live
GitHub: https://github.com/tejassudsfp/s...
It's experimental. The fundamentals are solid. Would love to hear what you're building and where this could help.