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

DAGent. Evaluate-then-Grow Planning for Deep Research Agents. By Hanwen Liu and two others, from New York University and New York University Shanghai. Published on the thirtieth of September. Its code repository is public but slight: six source files.

Deep research asks an agent to move through a large space of knowledge, pull evidence together from many sources, and change its plan as findings come in. Multi-agent systems built on a directed acyclic graph suit that work well. They can run sub-tasks in parallel, and each sub-task sits inside its own focused dependency context. The trouble lies in when the graph gets drawn. Existing graph-based agents lay out a plan for the whole task before anything runs, and they repair the graph only once a failure or a gap in the evidence shows up. The authors call this Plan-then-Patch, and they argue it is brittle for deep research for a simple reason. The system commits hardest at the moment it knows least, at the very start, and the revisions that follow spend computation on branches that should never have been planned at all.

DAGent turns the order around. Its planning is Evaluate-then-Grow: rather than drawing the whole graph up front, an Orchestrator grows the task graph one batch at a time. Each expansion is conditioned on confidence and uncertainty signals from the nodes that have already completed. The plan, in other words, is extended only as far as the evidence gathered so far can support.

Two further pieces sit around that loop. The first is a hierarchical context layer. By default it passes compact QueryDocs from node to node, while keeping the full execution traces available to be recalled on demand. The second concerns training. Because the graph's topology is recorded as it grows, it offers structural reinforcement learning signals that a recipe scoring only the final outcome cannot define. DAGRPO, an adaptation of GRPO, uses them in two places. It injects credit conditioned on the topology into the rollouts of the Executor, and it applies a structural compliance regularization to the plans of the Orchestrator.

The evaluation covers three benchmarks: BrowseComp-Plus, GAIA, and xbench-DeepSearch. With Qwen3 at the two hundred and thirty-five billion parameter scale, DAGent beats the strongest open-source baseline by five point three points on the first, five point eight on the second, and two points on the third.

That lead holds across four open-source backbones, and it extends to GPT-5 with three hundred and twenty-seven thousand tokens of context. Training gets its own test at the eight billion parameter scale, where DAGRPO beats an outcome-only GRPO baseline given the same budget by three points of average Pass@1. A final comparison keeps the architecture fixed and changes only the planning. Evidence-conditioned planning reaches higher accuracy than its Plan-then-Patch counterpart, while using fewer tokens, fewer tool calls, and fewer steps per task.

Underneath the benchmarks is an argument about timing. A plan drawn before any evidence exists is a guess, and patching a guess is costly. Growing the graph in step with what the completed nodes report keeps commitment matched to confidence. A recorded graph then pays a second dividend: its shape becomes a training signal in its own right, which a bare final answer could never supply.