Research
We publish across reasoning, tool-use, agent memory, and evaluation. Our work appears at ICML, NeurIPS, and arXiv.
Paper • arXiv 2026
We introduce a hierarchical memory architecture that achieves 94% retention accuracy across heterogeneous agent sessions. Our approach combines episodic memory buffers with compressed working memory representations, enabling agents to recall and apply knowledge from prior sessions without catastrophic forgetting. Evaluated on 10k+ agent trajectories across code engineering tasks.
Paper • ICML 2026
We show that structured tool-use during chain-of-thought reasoning improves accuracy by 37% on complex software engineering benchmarks. Our method interleaves natural language reasoning with API calls to compilers, linters, test runners, and documentation retrievers within a unified thought loop.
Model • 2026
Our flagship foundation model achieves state-of-the-art results on SWE-Bench, Tool-Use, and Multi-Step Reasoning benchmarks. Nexus-1 is trained on 3M+ trajectories of tool-mediated problem solving using a novel two-stage curriculum: supervised fine-tuning on expert demonstrations followed by RL from tool-use feedback.
Whitepaper • 2025
Why the next generation of AI requires a fundamentally new runtime. We analyze the limitations of chat-based interfaces layered on legacy systems and propose a set of architectural principles for building infrastructure designed for autonomous agents from the ground up.
Blog • 2026
A systematic methodology for measuring and improving the reliability of autonomous AI agents in production. We introduce coverage metrics, failure mode taxonomies, and a continuous evaluation pipeline that runs against every deployment.
Blog • 2026
Engineering lessons from deploying persistent memory across 10k+ agent sessions in production. We cover memory compaction strategies, retrieval latency, conflict resolution, and the surprising failure modes of long-lived agents.
Workshop • 2026
We extend persistent memory to multi-agent settings, showing that shared memory graphs enable teams of agents to coordinate, delegate, and resolve conflicts without centralized orchestration.
Preprint • 2026
A method for enforcing operational constraints during agentic decoding, guaranteeing that generated actions satisfy pre-defined safety policies without requiring post-hoc filtering.