A lightweight router for AI-agent handoffs selected either a typed dependency graph or natural-language instructions for each delegation, according to an arXiv preprint. The reported routed system matched or exceeded natural-language-only delegation on the main benchmark comparisons. The handoff harness used a separate executor and isolated the handoff format as the only variable.
Routed scored 47.3% on BrowseComp, 24.7% on τ-retail, 75.4% on BFCL and 51.7% on AppWorld. Those values matched or exceeded natural-language-only delegation in each comparison.
A small router, two handoff formats
The router used the task pattern to choose between the two formats. It cost about 155 tokens and added 0.15% overhead. A graph-formatted delegation was a typed DAG, or a one-way map of dependencies, with eight node types and seven edge relations. The other path used natural-language prose.
On τ-retail, the graph-only setup was 12.7 percentage points higher than the natural-language setup in 150 paired trials, with p<0.01. On BrowseComp, the reported difference was 8.7 points, with a 95% confidence interval from 2.7 to 14.7 points and p<0.05.
AppWorld produced a different comparison
On AppWorld, graph-only handoffs were 14.6 percentage points lower than the natural-language comparison. The reported 95% confidence interval ran from 22.8 points below to 6.4 points below. Routed scored 51.7% on the benchmark in the main results.
The reported token accounting
The reported efficiency figures included average handoff compression of 2.1 times and router overhead of 0.15%. In the full τ-retail accounting, Routed used 461 tokens versus 730 for natural-language delegation, a 1.6-times compression. The token counts were measured, while latency figures in the same accounting were estimated.
Additional comparisons
An additional τ-airline ablation covered 150 trials. Graph-only delegation was 4.0 percentage points lower, while the router sent 98% of tasks to natural language. The routed condition recovered parity.
A broader protocol comparison used 50 τ-retail tasks with three seeds, producing 150 trials per protocol. All schema-aware methods outperformed schema-unaware methods in that comparison, and the routed graph protocol was the only zero-training protocol among the top five.
An AppWorld oracle analysis compared the routed result with a per-task choice of the better format. Routed scored 51.7%, compared with 60.3% for the oracle, an 8.6-point gap. The reported decomposition attributed 5.2 points to the graph format and 3.4 points to natural language. The oracle is an upper-bound analysis rather than a deployable intervention.
The test depended on the executor
The reported graph comparison included instructions for the receiving agent. Passing the same JSON graph to a standard executor prompt produced no gain, while the graph setup used graph-aware executor guidance.
The main evaluation contained 1,052 trajectories. Reported confidence intervals used paired bootstrap estimates at the 95% level.
An automated error taxonomy applied to 345 τ-bench trajectories produced a reported 76% inter-agent-misalignment figure.
A second orchestrator
With GPT-5 mini as the orchestrator, Routed was also higher than natural-language delegation in the reported reruns. The scores were 68% versus 65% on BrowseComp, 85% versus 82% on BFCL, and 52% versus 50% on AppWorld. The reported check was limited to those rerun families.
The result has a narrow reach
The evidence remains tied to the reported handoff harness and its benchmark trajectories. The harness isolated handoff format as the only variable, and graph execution used executor guidance as part of the setup. The reported scores should be read as results from that setting.
The router selected formats by task pattern, while the AppWorld oracle was an upper-bound analysis rather than a deployable intervention. The reported material therefore covers task-pattern selection and an oracle comparison, not a fine-grained or mid-trajectory routing system.
Paper data and sources
Original title: Routed Graph Handoff: Adaptive Format Selection for Multi-Agent LLM Delegation
Authors: Pratyay Banerjee, Ankit Chadha
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text