AgentLens: Production-Assessed Trajectory Reviews for Coding Agent Evaluation
Instead of judging coding agents with a single pass/fail bit, AgentLens evaluates the full “trajectory” – instruction following, tool use, self-verification, recovery from mistakes, and the explanations given along the way. The benchmark pairs formal checks with LLM-written trajectory reviews, making it practical for nightly regression testing and for diagnosing why an agent’s behavior changed.
QANTIS: Hardware-Calibrated Sequential POMDP Belief Updates on IBM Heron
This work treats a quantum processor as a calibrated service for updating beliefs under partial observability, then tests whether the returned posterior remains stable for downstream classical planning. Using controlled hardware experiments on IBM Heron, the authors show that specific amplification and boundary-aware estimation strategies can preserve the action selection relative to exact Bayes posteriors across longer sequential horizons.
When Does In-Context Search Help? A Sampling-Complexity Theory of Reflection-Driven Reasoning
The paper analyzes in-context search (iteratively generating, critiquing, and revising reasoning) as approximate inference over reasoning traces, focusing on how many sequential attempts are needed for high success. It predicts exponential gains when reflections reliably pinpoint early mistakes, but shows those benefits vanish when reflection quality fails – and it argues the gains are learnable and consistent with stagewise RL formulations with verifiable rewards.
ChatGPT Work is an agent that can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work
OpenAI’s ChatGPT Work positions ChatGPT as an action-taking agent that can stay with multi-hour projects and convert a user’s goal into completed work, powered by the GPT-5.6 model suite. The move extends the “chat” experience toward a more Codex-like workflow where tasks aren’t just discussed – they’re executed across apps and documents.
Evaluating SageMath-Augmented LLM Agents for Computational and Experimental Mathematics
This paper builds a ReAct-style agent loop that combines LLM reasoning with verifiable feedback from SageMath, using updated documentation support via Context7. On RealMath-style research tasks, SageMath access boosts performance across frontier models and the authors propose multi-step post-processing and validation to improve benchmark reliability.
Grounding Spatial Relations in a Compact World Model: Instruction Leakage and a Goal-Free Dynamics Fix
The study shows a subtle failure mode in goal-conditioned world models: high “relation accuracy” can come from instruction transcription rather than true perception and grounding. The authors identify instruction leakage when the scored quantity is recoverable from the instruction, then propose a fix – keep the goal out of the dynamics and supervise the read path – restoring instruction-independent grounding.
Meta says its new AI model is ready to compete on coding
Meta is opening access to its Muse Spark 1.1 model through a Meta Model API, targeting coding workflows and more advanced bug detection and fixing. The report highlights improved support for agentic, multi-app and multi-agent workflows plus native multimodal perception across images, videos, and documents.
Cost-Effective Agent Harnesses for Abstract Reasoning and Generalization on ARC-AGI-1
Instead of relying on heavy test-time compute or benchmark-specific fine-tuning, the authors study what an open-weight model can recover using architecture alone. Their Explorer-Definer Pipeline plus a Reflective Orchestrator achieves strong ARC-AGI-1 results under a strict budget, and the paper argues improvements are generation-bound by analyzing unbiased pass@k effects.
GPT-5.6 is now the preferred model in Microsoft 365 Copilot
Microsoft says GPT-5.6 is now its preferred model for Microsoft 365 Copilot, aiming for stronger capability across Word, Excel, PowerPoint, Chat, and Cowork. The update frames GPT-5.6 as delivering better performance per dollar and more ability on demand for high-impact work.
GPT-5.6: Frontier intelligence that scales with your ambition
OpenAI describes GPT-5.6 as frontier intelligence designed to scale with the complexity of a user’s goals, pairing it with a broader suite of model variants. The announcement emphasizes capability gains across ambitious, multi-step work rather than just incremental improvements on smaller prompts.
The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI
This paper argues the biggest lever against “token maxing” isn’t just model choice – it’s the agent harness, the orchestration layer that manages context, tools, turns, delegation, and governance. In a controlled evaluation swapping only orchestration, the harness cuts tokens per task and blended cost significantly while keeping completion quality roughly on par, introducing the idea of “harness leverage” where better baselines benefit most.
Learning social norms enhances compatibility in dynamic human-AI coordination
Instead of aligning agents to demonstrations alone, the authors extract quantifiable social-norm principles from thousands of human interactions and test them in closed-loop coordination scenarios. In a pedestrian-vehicle style task, norm-informed LLM behavior boosts coordination scores dramatically compared to a baseline and can even outperform human-human interactions.
OpenAI rolls out GPT-5.6 after government greenlight — and announces ‘ChatGPT Work’
According to the report, OpenAI received government approval for a public rollout of GPT-5.6 after a limited preview, with the company calling it the best model it has produced. The same day, OpenAI introduced ChatGPT Work as a new agent that blends ChatGPT and Codex-style execution so non-technical users can delegate more ambitious goals.
FL Studio 2026 turns its AI chatbot into your assistant engineer
Image Line upgrades its Gopher assistant in FL Studio from a “glorified instruction manual” into an agent that can perform actions – including laying down drum patterns and applying effects like gated reverb when prompted. The hands-on reporting frames it as assistant-engineer behavior inside a creative workflow, with remaining limitations around automation drawing and some types of music editing.