Agentic AI

Occamy-1.0 is a cost-efficient model built on the Qwen3.6-35B-A3B checkpoint, designed specifically for co-work agents that execute long-horizon workflows. It leverages execution-grounded data and replayable trajectories to optimize for state tracking, coordination, and recovery rather than just peak reasoning. The model uses staged post-training to balance capability with the latency and cost constraints of multi-step agent episodes.

• Optimized for multi-step agent workflows, not just single-turn reasoning tasks.
• Built on Qwen3.6-35B-A3B, focusing on efficiency and state management.
• Uses execution-grounded data to improve coordination and recovery skills.
• Staged post-training targets the Pareto frontier of cost vs. capability.

A new paper introduces GAUGE, an offline protocol evaluating the validity of persona-driven LLM user-simulators as judges for task-oriented agents. Testing 25 agents across six providers on the tau-2-bench and SimulatorArena benchmarks reveals a critical satisfaction-success gap. The study finds that high user satisfaction ratings provide essentially no information about whether the agent actually completed the task successfully.

• LLM judge satisfaction scores are poor proxies for actual task completion success.
• GAUGE protocol separates ranking validity from construct validity in agent evaluation.
• Blind panel ratings of satisfaction do not correlate with verifiable rewards.
• Offline simulation gates may promote agents that are polite but ineffective.

AMD introduces AMDKernelVault, an open dataset and training framework specifically for optimizing kernels on recent CDNA GPUs. The project addresses the NVIDIA-centric bias in current LLM-based kernel agents by providing HIP and Triton corpora with 62,153 verified HIP samples and nearly 40,000 Triton kernels. It features agent-driven pipelines, HIPKernelGen and TritonKernelGen, which convert PyTorch references into optimized kernels validated under ROCm. The team also fine-tunes Qwen3-8B using supervised learning and execution-aware reinforcement learning to improve agent performance.

• Fills the gap for AMD-specific LLM kernel agents, moving beyond CUDA/NVIDIA bias.
• Provides a large, execution-verified corpus of 62k+ HIP and 39k+ Triton kernels.
• Uses agentic pipelines to transform PyTorch code into ROCm-validated HIP/Triton kernels.
• Fine-tunes Qwen3-8B with execution-aware RL to enhance kernel generation quality.

A new arXiv paper isolates the 'harness effect' by testing identical large language models against different agentic toolkits using a private, contamination-controlled suite of 256 tasks. The study ran paired contrasts, such as Claude Opus 4.8 with both the native SDK and DeepAgents, and GPT-5.5 with OpenAI Codex and DeepAgents, grading 792 runs via an isolated oracle. Results indicate that switching harnesses does not yield a statistically significant average advantage in task resolution compared to vendor-native configurations.

• Vendor-native model-harness pairings are not inherently superior to third-party or open harnesses like DeepAgents.
• Performance gains from agentic systems may stem from the model itself rather than the surrounding tooling infrastructure.
• The study uses a contamination-controlled private suite, mitigating data leakage concerns common in public benchmarks.
• Practitioners should evaluate harness flexibility and integration costs rather than assuming vendor lock-in improves outcomes.

WinSyn addresses the gap between existing benchmarks and real-world enterprise complexity by generating synthetic datasets based on realistic workplace scenarios. The pipeline creates synthetic emails to support Retrieval-Augmented Generation and Deep Research agents, providing both long-form and short-form questions with gold answers. This approach aims to capture the challenges of information spread across evolving and potentially conflicting enterprise artifacts.

• Generates synthetic enterprise emails to simulate complex, realistic QA scenarios.
• Provides both long- and short-form questions with gold answers for evaluation.
• Targets limitations of current benchmarks that lack real-world data complexity.
• Supports testing of RAG and Deep Research agents in enterprise settings.

The AIM framework addresses the limitation of LLMs being confined to single-session contexts by enabling persistent memory management across multi-agent, multi-user environments. It dynamically classifies information into private scopes for individual users and public scopes accessible to all, allowing shared knowledge to improve downstream responses. This approach ensures that user preferences evolve over time while maintaining strict privacy boundaries between personal and community data.

• Shifts LLM memory from ephemeral sessions to persistent, cross-user knowledge bases.
• Dynamically enforces privacy by separating private user data from public shared memory.
• Enables multi-agent systems to leverage collective insights without compromising individual privacy.