AutoRAS proposes a framework for automatically designing robust agentic systems by generating sequences of symbolic primitives that encode both structural connectivity and behavioral actions. It optimizes these sequences using safety signals from execution and flow-based objectives, achieving superior performance in both normal and adversarial conditions with minimal degradation under attacks.
CORTIS enables task-oriented voice agents to generate structured speech outputs by fine-tuning spoken language models using only text-form task supervision. It outperforms ASR-LLM cascades under acoustic degradation, especially in preserving high-level task semantics, without requiring paired speech-target annotations during training.
w$^{2}$VLA introduces a modular vision-language-action model that decouples declarative and procedural knowledge. By restructuring information flow, it enables robust behavior cloning and zero-shot skill transfer to novel, dissimilar objects.
A new approach verifies agentic AI workflows during design by modeling them as compositions of reusable building blocks. It applies twelve structural rules to ensure compatibility, reliably detecting design flaws even after structural transformations like task splitting.
A unified zero-shot framework, ZeProM, uses a pre-trained Video-Language Model to jointly perform procedural mistake detection and temporal action segmentation. It achieves up to 4.4 point improvement in EDA and 2.0 point in F1@.5 on EgoPER tasks, matching or exceeding supervised methods without task-specific training.
MiniMax 2.7, a 47 tera-parameter model, operates on a 96GB VRAM system with 192GB DDR5 RAM using an MSI B840 board and 9900X CPU. It runs as an agent-class model with strong instruction following and tool calling, supported by a round-robin loop with three CPU-based sequencing agents and a dense 12B model that monitors for errors.
Claude v2.1.187 introduces sandbox credentials blocking, org-configured model restrictions, mouse click support in fullscreen, and fixes for command failures, tool hangs, and UI stability. Updates also improve structured output handling, agent depth tracking, and plugin management, with enhancements to VSCode and terminal compatibility.
Tmax-27B is a terminal agent based on Qwen3.6-27B, trained with DPPO (RL), achieving 43% on Terminal Bench 2.0 and 69% on TB Lite. To run on consumer GPUs, it is quantized using importance-matrix-calibrated GGUF models from 2 to 5 bits per weight, with a grafted MTP head enabling speculative decoding. IQ2_XS at 8.5 GiB achieves 70% pass rate in agentic coding tasks, outperforming plain quantization and demonstrating stable tool-call generation.
Claude Tag allows teams to tag @Claude in Slack to delegate tasks, with access to selected channels, tools, and codebases. It learns from channel context, works asynchronously, and takes initiative by proactively updating users on relevant information. Today, 65% of Anthropic’s product team code is created by internal Claude Tag, and it’s now available in beta for Claude Enterprise and Team customers.
Hayden has developed the knot harness to manage long-running local LLM tasks. It enables reusable workflows with agent profiles, file system event monitoring, and automatic triggers, using Pi.dev as the default agent.
The creator of EverFern asks which local models work best for agentic workflows and browser/computer use. They note that model intelligence is rarely the bottleneck, with reliability and recovery systems being more critical than model choice.
A user asks whether supervised fine-tuning (SFT) followed by reinforcement learning (RL) is still recommended for training Qwen 3.5 4B or 9B agents for multi-tool use, or if RL-only approaches yield better results. The post also seeks guidance on reward design and handling parallel tool execution in agent workflows.
Group-Graph Policy Optimization (G2PO) introduces a graph-based approach to enhance long-horizon agentic reinforcement learning by transforming interaction trajectories into state-transition graphs. It enables group-aggregated state-value estimation and edge-centric advantage calculation, improving credit assignment and reducing variance, and achieves up to 22.2% success rate improvement over GRPO on WebShop, ALFWorld, and AppWorld benchmarks.
PhoneBuddy combines real and mock app environments to train open models for phone use. It improves task success rates from 36.67% to 45.33% on real phones and from 60.3% to 83.2% on AndroidWorld, showing mock-app training complements but does not replace real-app RL.
ToolGraph enhances multi-turn tool-using agents by integrating schema topology, transition weights, and history-aware controls. Training with DPO on 161 divergence-point preference pairs improves performance: ToolGraph+DPO achieves a 16.8% relative reward gain over baseline, especially in airline and retail tasks, with reward positivity emerging as the key diagnostic signal.
AFTER introduces a benchmark of 382 enterprise tasks across six roles and 22 skills to assess skill transfer across tasks, roles, and models. Results show procedural memory improves performance by 3.7-6.7 points per refinement and achieves 73.1% cross-model accuracy, with some skills generalizing broadly and others specializing to role-specific workflows.
CUGA introduces a lightweight harness enabling developers to build real agentic applications. It includes 24 working examples demonstrating practical implementations across various use cases.
AgentCIBench introduces a benchmark to assess privacy risks in computer-use agents. It identifies three key failure modes—visual co-location, task-ambiguity overshare, and recipient misalignment—and finds that 11 of 15 evaluated agents leak personal data in over 50% of scenarios, with an average leakage of 67.9%.
Tmax presents the strongest open RL recipe for terminal agents, achieving 27% on Terminal-Bench 2.0 with only 9B parameters. It uses a novel data taxonomy to generate over 2.5x more terminal environments than prior datasets, enabling efficient training with a simple, outcome-only recipe. The dataset, models, and code are open-sourced at https://github.com/hamishivi/tmax.
SelfCompact enables language models to autonomously decide when and how to compact accumulated context during reasoning. By combining a model-invoked summarization tool with a lightweight rubric that guides compaction based on trajectory structure, it achieves effective adaptive compaction without fine-tuning. Results show it matches or exceeds fixed-interval methods on math and agentic search benchmarks, improving baselines by up to 18.1 points on math and 5-9 points on search, at 30-70% lower token cost.