Hi HN,

I'm Marco — ex-ethologist turned AI systems engineer — and I built something I needed but couldn't find: an open-source framework to wire *cognition like circuits* — not spaghetti prompt chains.

It's called *OrKa*: the *Orchestrator Kit for Agents*.

WHY I BUILT IT • Tired of black-box LLM chains (LangChain, AutoGPT, etc.) • Needed a way to fork/join reasoning paths • Wanted true traceability, memory, and reproducibility • Inspired by how cognition decays, splits, reconverges • Built from scratch with Redis Streams, YAML logic, and local model support

I wanted reasoning I could *see*, *debug*, *version*, and *replay* — like functional circuits.

WHAT ORKA IS

> YAML-defined cognition graphs (versioned, inspectable)

> Fork/Join execution with trace replay

> Router agents w/ confidence-weighted branching

> Redis or Kafka backends

> Scoped memory (episodic, procedural, etc.)

> Visual UI (React-based): https://orka-ui.web.app

> ServiceNodes: RAG, Memory fetch, Writer, Embedder

> Full local+remote LLM support via LiteLLM / OpenAI / Ollama

> 76%+ test coverage, deterministic behavior

BENCHMARK (REAL)

• 1000 orchestration runs (2-agent pipeline)

• DeepSeek-R1 (1.5B) via Ollama on Pop!_OS

• Avg latency: ~7.6s per agent

• Zero agent drift across runs

• Total cost (simulated): ~$0.49

• CPU temp: stable 88–89°C

• RAM: < 5.3 GB

• RESULTS: https://github.com/marcosomma/orka-reasoning/blob/master/docs/orka_linux_stressTest_toShate.zip

LINKS

• PyPI → https://pypi.org/project/orka-reasoning/

• GitHub → https://github.com/marcosomma/orka-reasoning

• Examples → https://github.com/marcosomma/orka- reasoning/tree/master/examples

• UI (Docker)→ https://hub.docker.com/r/marcosomma/orka-ui

FEEDBACK WELCOME

• Is this the right abstraction layer for agentic reasoning?

• Should this remain infra-first or move toward hosted cognition-as-a-service?

• Anyone else frustrated with current LLM toolchains?

Thanks for reading. AMA. _