I revived Shandu, an open-source multi-agent research system focused on reproducible outputs instead of chat-style summaries.
It uses a lead orchestrator that runs iterative research loops, parallel subagents for search/scrape/extract, and a citation agent that builds/normalizes the final reference ledger.
You get both a Rich CLI control deck and a Gradio GUI with live telemetry, task traces, citation tables, cost coverage, and one-click markdown export/download.
Core ideas:
- iterative planning + synthesis instead of one-shot prompting
- explicit evidence records + normalized numeric citations
- model/provider flexibility via Blackgeorge/LiteLLM
- SQLite-backed run/memory tracking for inspectability
I revived Shandu, an open-source multi-agent research system focused on reproducible outputs instead of chat-style summaries.
It uses a lead orchestrator that runs iterative research loops, parallel subagents for search/scrape/extract, and a citation agent that builds/normalizes the final reference ledger. You get both a Rich CLI control deck and a Gradio GUI with live telemetry, task traces, citation tables, cost coverage, and one-click markdown export/download.
Core ideas:
- iterative planning + synthesis instead of one-shot prompting - explicit evidence records + normalized numeric citations - model/provider flexibility via Blackgeorge/LiteLLM - SQLite-backed run/memory tracking for inspectability