Show HN: SkillGraph – Open-source agentic framework with skills instead of tools (github.com)

SkillGraph is an open-source, experimental agent framework that replaces the common “tool-calling” pattern with autonomous “Skills” (subagents) that take end-to-end responsibility for specific tasks. Instead of a central agent planning which low-level tools to call, the agent routes requests to domain skills that orchestrate tools, maintain their own prompts and state, and support multi-turn interactive workflows. The approach is designed to cut token usage, improve reliability and error recovery, and make complex, multi-step conversations and business logic easier to implement and control. Technically, SkillGraph combines Subject/Object state tracking, a fast Utility LLM (Llama-3-8B ~200ms) with an Anthropic model fallback chain (Sonnet 4.5 alpha, Haiku 4.5 beta), and aggressive caching (Redis + Anthropic prompt caching) that claims up to 89% system-prompt savings. It uses Redis (<5ms) with PostgreSQL fallback, context compression every 10 messages, pgvector semantic search for historic context, and a multi-layer moderation pipeline. Features include async-first streaming, parallel skill execution, interactive skill modes, Prometheus/Grafana metrics, Kubernetes configs, and multi-provider support (Anthropic, OpenAI, Bedrock, Hugging Face, etc.). Performance targets: simple queries <200ms, complex 0.5–2s, cached <50ms. It’s functional but experimental—promising for teams seeking lower cost, better control, and robust multi-turn behavior, but not yet battle-tested at large scale.