{"author":"NeuralOperator","children":[{"author":"bettaher_adam","children":[],"created_at":"2026-03-29T01:29:32.000Z","created_at_i":1774747772,"id":47559657,"options":[],"parent_id":47510639,"points":null,"story_id":47510639,"text":"The skeptic loop concept is the most interesting part here. \nOne thing worth considering for the validation layer: \nsigning the intermediate outputs between agents with HMAC \nso you can trace exactly which agent in the chain produced \nwhich conclusion. Makes debugging multi-agent pipelines \nsignificantly easier when something breaks mid-chain.

Also curious how you handle the case where Supervisor \ngets conflicting conclusions from two specialized agents \u2014 \ndoes it default to the skeptic or does it re-run?","title":null,"type":"comment","url":null}],"created_at":"2026-03-24T22:48:42.000Z","created_at_i":1774392522,"id":47510639,"options":[],"parent_id":null,"points":1,"story_id":47510639,"text":"Hi HN, I built EvidionAI \u2014 an autonomous research pipeline where a Supervisor \norchestrates a loop of specialized agents to answer scientific questions \nend-to-end.

The workflow:\nSupervisor \u2192 Search (DDG + arXiv + Wikipedia) \u2192 Code (Python in sandboxed Docker) \n\u2192 Analysis \u2192 Skeptic \u2192 back to Supervisor if the conclusions don't hold up.

The focus is not just on execution, but on validation \u2014 the system actively \ntries to break its own conclusions via a skeptic loop.

Stack: LangGraph, LangChain, FastAPI, ChromaDB, SQLite (before that, there was PostgreSQL and Redis), nginx. \nWorks with Ollama (local/cloud) and any OpenAI-compatible API. \nOne-command Docker Compose setup.

https://github.com/Evidion-AI/EvidionAI

I'm open-sourcing this because a lot of similar projects are emerging right now, \neach exploring different approaches to autonomous research and agent orchestration.

I'm building this solo, and it seems more valuable to share the approach early \nrather than develop it in isolation \u2014 especially since the space is evolving fast \nand there\u2019s clearly convergent interest in this direction.

Would really appreciate feedback on the agent architecture \u2014 especially the \nSupervisor routing logic, which is currently the most fragile and interesting part.","title":"EvidionAI \u2013 open-source multi-agent research system built on LangGraph","type":"story","url":null}