Megadose The AI curve, compressed.

DAGent: Evaluate-then-Grow Planning for Deep Research Agents

· HF Daily Papers ·
DAGent delays graph expansion until completed research nodes report confidence and uncertainty.

The paper argues that fixed upfront DAG plans are brittle for deep research because they commit before enough evidence exists. DAGent instead grows the task graph in batches through an Orchestrator, while a context layer passes compact QueryDocs and keeps full traces available for recall. The authors also introduce DAGRPO, a reinforcement-learning variant that uses DAG structure for credit assignment and plan regularization. They report gains over open-source baselines across BrowseComp-Plus, GAIA, and xbench-DeepSearch, with additional results at Qwen3-8B and GPT-5 scale. HF Daily Papers' note

score 4

Categories: Research