Multi-agent systems are now routine, but one decision goes unmeasured: when an orchestrator hands a subtask to a peer model, does it pick the right one? We built DecisionBench to score the handoff itself, not just the final answer, across an 11-model, 7-vendor pool over 23,375 task instances. The headline finding: there is enormous performance left on the table. A perfect-delegation ceiling sits 15 to 31 points above where today's systems land, on every benchmark we ran. That gap is the prize, and skill-aware orchestration is how you capture it. Read more about it in our blog post: openmesh.ai/news/decisionben…

Great day at ACM Conference on AI and Agentic Systems @CAISconf Ion Stoica’s @istoica05 talk on efficient search was a highlight: stagnate -> explore, improve -> exploit. And what a response at our AgentPulse and IntelligenceArena posters, two sessions, both packed. Thank you to everyone who came by. Continuous eval for agents in deployment is striking a nerve, and the conversations proved it.

We will be at the ACM Conference on AI and Agentic Systems (CAIS) tomorrow @CAISconf Come find the OpenMesh team at our poster station. We’d love to talk about your agentic workflows. Posters: “AgentPulse” “IntelligenceArena”

As LLMs become increasingly powerful, the ability to effectively delegate and collaborate with other models may itself become a critical capability for long-horizon agentic systems. It was a lot of fun working with these passionate young researchers on DecisionBench! @MichaelYGao @meganwxng @YYL_GB

Final-task accuracy can hide a broken multi-agent system. Your agent may “solve” the task while making bad delegation decisions the whole way through. That’s why we built DecisionBench: an eval for whether agents delegate well, not just whether the final answer looks right. 🧵

Excited to share what we’ve been working on @OpenMeshAI DecisionBench aims to measure how well LLM agents emergently delegate sub-tasks to peer models during long-horizon workflows, using process-level metrics that quality-only evaluation misses.

Your multi-agent system might be delegating to the wrong model on every single subtask, and your eval would never tell you 🧵 Introducing DecisionBench: the first benchmark that measures whether agents actually delegate well, not just whether the task got done. arxiv.org/abs/2605.19099

Three strong signals today: DecisionBench (delegation quality), UCCI (calibrated cascade routing), and Gemini 3.5 Flash (agentic speed). Common thread: advantage is moving from model IQ to orchestration policy—who handles which step, at what confidence, at what cost.

DecisionBench's methodology creates emergent delegation patterns: 1) Peer model pool with known capabilities (GPT-4, Claude, Gemini, etc.) 2) call_model interface optional read_profile channel 3) Skills deterministically annotated across all models 4) Tasks from GAIA, tau-bench, BFCL

The promise of autonomous AI agents often rests on their ability to reason through complex, long-horizon tasks. Yet, we frequently overlook the "delegation problem", which is the moment an agent decides it needs to offload a sub-task to another model or tool. Enter DecisionBench, a new benchmark designed to stress-test how these systems handle emergent delegation. This is not just a technical curiosity, but a governance challenge. When an agent autonomously decides to delegate a critical fact-checking or drafting process to an external module, where does the chain of accountability end? DecisionBench forces us to confront the "black box" of agentic decision-making, moving us beyond simple benchmarks toward a more rigorous understanding of agency, error propagation, and the risks of unchecked automation in high-stakes environments. If we intend to integrate these workflows into the practice of law, we must ensure our agents possess not just intelligence, but a sense of jurisdictional competence. Source: DecisionBench: A Benchmark for Emergent Delegation in Long-Horizon Agentic Workflows, by Yuxuan Gao, Megan Wang, Yi Ling Yu, Zijian Carl Ma, Ao Qu

Long-horizon AI agents need to delegate subtasks to specialized models, but nobody's been measuring this systematically. DecisionBench fixes this: 11 models, 3 task suites, deterministic skill annotations, and multi-axis evaluation covering quality, cost, latency, and routing fidelity.

AgentPulse: A Continuous Multi-Signal Framework for Evaluating AI Agents in Deployment Yuxuan Gao, Megan Wang, Yi Ling Yu arxiv.org/abs/2604.24038 [𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙲𝙻 𝚌𝚜.𝚂𝙴]