Astraea - Build deterministic, multi-hop reasoning agents with Astraea

Traditional frameworks force a "push" context model where every node receives the complete state. When a simple triage agent or supervisor agent runs, they ingest massive database logs, racking up excessive token costs across entire swarms. Astraea implements a "pull" context model. The router and supervisor agents remain extremely token-light by querying simple metadata, while only specialist agents pull massive payloads. This yields up to a 3.3x reduction in cumulative input tokens.