top of page

Research on the Next Generation of Agentic Infrastructure

Benchmarking, evaluating, and optimizing real-world AI systems across cost, latency, and quality

Product-driven Research. Research-driven Product.

We believe the next wave of AI progress will come not only from stronger models, but from better systems around those models: how inference is deployed, how workflows are routed, how outputs are grounded in external information, and how performance is measured in production. OpenMesh is built around that view. Our approach is to treat product and research as mutually reinforcing. Real-world usage exposes failure modes, performance bottlenecks, and cost-quality tradeoffs that do not appear in isolated benchmarks. Research helps us convert those observations into better architectures, stronger routing policies, and more reliable production systems.

OpenMesh: A Lab for Production AI Systems

May 17, 2026

Distribution-Free Uncertainty for Continuous Agent Evaluation

Accepted to International Conference in Machine Learning (ICML) 2026 AgenticUQ Workshop

May 7, 2026

DecisionBench: Benchmarking Skill-Aware Emergent Orchestration in Long-Horizon Agentic Workflows

In Submission to NeurIPS 2026

Token Arena: A Continuous Benchmark Unifying Energy and Cognition in AI Inference

Public benchmarks for AI inference compare systems at the level of the model or the provider, yet the unit at which deployment decisions are actually made is finer: the endpoint, defined as the (provider, model, SKU) tuple at which a particular quantization, decoding strategy, region, and serving stack are exposed. Two endpoints serving nominally the same model can differ substantially in latency, price, energy footprint, and even output quality, and these differences are largely invisible in current leaderboards. Token Arena is a continuous benchmark that measures inference at endpoint granularity along five core axes (output speed, time to first token, workload-blended price, effective context, and quality on the live endpoint) and synthesizes them, together with a modeled energy estimate.

AgentPulse: A Continuous Multi-Signal Framework for Evaluating AI Agents in Deployment

AgentPulse is a continuous, multi-signal framework for evaluating AI agents in deployment, scoring 50 agents across 10 workload categories along four factors — Benchmark Performance, Adoption Signals, Community Sentiment, and Ecosystem Health — aggregated from 18 real-time signals across GitHub, package registries, IDE marketplaces, social platforms, and benchmark leaderboards. The core motivation is that static benchmarks measure what agents can do at a fixed point in time but not how they are adopted, maintained, or experienced in practice; across the full registry, benchmark scores and adoption are nearly uncorrelated (ρ=0.05), confirming that task-completion capability does not predict which tools developers actually choose.

Accepted to Conference on AI and Agentic Systems (CAIS) 2026 AID-Wild Workshop

IntelligenceArena: A Quantitative Framework for Real-Time Scoring of Frontier LLMs and AI Agents

Intelligence Arena introduces a quantitative framework for evaluating frontier models based not only on raw capability, but on efficiency. Rather than asking which model performs best in isolation, this work measures intelligence per token across diverse tasks, capturing the tradeoffs between quality, cost, and latency in real-world usage. By combining multi-factor evaluation signals, Intelligence Arena provides a more practical and production-aligned view of model performance, enabling developers and enterprises to identify the most effective models under constrained budgets and dynamic conditions.

Accepted to Conference on AI and Agentic Systems (CAIS) 2026 AgenticSE Workshop

Interested in researching with us?

We are interested in speaking with researchers, engineers, and early partners who care about the future of agentic infrastructure systems.

Contact Us

bottom of page