Agent-Field/af-deep-research

Implements iterative self-correcting research loops where parallel agents explore multiple angles, filter evidence against quality thresholds, identify gaps, and spawn targeted sub-queries until completion—orchestrating ~10,000 logical agent invocations per research. Built on the AgentField runtime, it returns structured outputs (typed entities, relationships, evidence, and cited documents) via a single REST API call with SSE streaming. Integrates with multiple search providers and LLM backends, supporting self-hosted deployment with local or air-gapped models.