What problem does it solve?
Autoresearch automates the repetitive, manual workflow of discovering missing concepts and sourcing authoritative material to expand a curated knowledge base, preventing endless manual searches and inconsistent ingests. It runs bounded, multi-round research loops that detect documentation gaps, fetch web sources, create immutable raw stubs, and optionally run the KB INGEST pipeline until no new concepts appear or a configured cap is reached.
Core Features & Use Cases
- Multi-round gap detection: Reads wiki index and related pages to produce explicit gap lists and diffs across rounds.
- Automated query formulation and parallel fetch: Generates precise search queries for each gap and performs WebSearch + WebFetch up to pages_per_round while honoring allowlists/denylists.
- Create-only raw stubs and safe INGEST: Writes new raw/ files (never modifies existing raw files), runs the KB's INGEST workflow when in wiki mode, and updates index, logs, and recently-added registers.
- Safety and provenance: Enforces reviewed=false, records contradictions, requires inbound links, and never promotes to hot cache; produces a final structured markdown report saved to outputs/.
- Use case: Expand the Agentic-KB on "supervisor-worker multi-agent patterns" for a literature-backed wiki section without manual round trips.
Quick Start
Run autoresearch on "supervisor-worker multi-agent patterns" with max_rounds=2 pages_per_round=4 and source_allowlist=[arxiv.org,anthropic.com] so the agent performs two bounded rounds and saves the final report to outputs/.
