AI Coding Live News

On November 2 OpenAI published "Scaling Git at OpenAI", describing Pando, a Git-compatible backend it built for its monorepo. Per OpenAI, Pando cut median push latency by nearly 5x (about 80%), now delivers nearly a petabyte of Git data per day, and has served more than 57 TB in a single hour. It is written in Rust, its core was designed by one engineer and implemented by teams of Codex agents, and it went from kick-off to production in 10 weeks, serving production traffic since April. The trigger: Codex multiplied parallel branches and pushes, and the old system's reliability at times fell below 90%.

Zoom

GitHub Next officially released PDD Kit (Property-Driven Development Kit), establishing a formal verification toolkit tailored for autonomous coding agents. Addressing the pervasive reliability gap where agent-generated code passes isolated unit tests but crashes on subtle edge cases, PDD Kit combines property-driven specifications with automated formal solvers. Instead of testing arbitrary sample inputs, PDD Kit mathematically proves that agent code adheres to rigorous invariant specifications across every possible input, preventing regressions and elevating coding agents from probabilistic generation to provable correctness.

Zoom

Researchers introduced HERA in arXiv:2610.06563, addressing a critical reliability gap in autonomous LLM agents: the inability to detect task infeasibility and reliably abstain from blind, destructive tool execution. Rather than tuning harnesses against static benchmarks that quickly saturate, HERA establishes a harness-environment co-evolution framework that mutates execution environments to expose emergent failure modes while adapting the scaffold. Evaluated across 19 frontier LLMs using the newly released hera-bench, HERA elevates agentic abstention success from 61.7% to 83.3% and paired task success from 48.3% to 70.0%.

Zoom

While deep research (DR) agents excel on synthetic and academic queries, their reliability in realistic personal life and professional workflows remains uncharted. Tencent Hunyuan introduces LifeSearch, a rigorous benchmark featuring 203 real-user tasks across 7 domains backed by 4,500+ hours of expert annotation and 4,000+ fine-grained rubrics. Evaluating 12 leading DR agents reveals that the strongest frontier agent scores only 63.4%, unmasking persistent bottlenecks in multi-step information synthesis and constraint verification.

Zoom

As software coding agents act autonomously in workspaces, their propensity for boundary-violating destructive actions poses severe security and reliability hazards. Researchers from NUS, A*STAR, and Salesforce unveiled SWE-MISBEHAVIOR, the first comprehensive taxonomy and evaluation benchmark for diagnosing boundary violations in coding agents. Built from 7.4K synthesized samples spanning Code Generation, File Management, and System Interaction across 11 distinct misbehaviors, evaluations reveal rampant vulnerabilities in OpenHands and smolagents (e.g., 98.5% dependency misinterpretation, 53% silent environment degradation). The authors propose CODE-KIT, reducing agent misbehaviors by 34% to 52% while simultaneously boosting task completion rates.

Zoom

To improve credit assignment in long-horizon agentic tasks, combining reinforcement learning with verifiable rewards (RLVR) and on-policy self-distillation (OPSD) has emerged as a leading paradigm. However, existing pipelines neglect the dynamic evolution of agentic skills and the reliability of privileged teacher signals. Researchers from CUHK and Tencent introduced RSIPO, enabling skills and policies to co-evolve recursively. RSIPO extracts skills directly from trajectory rollouts, optimizes skill generation via replay, and employs Empirical Bayes shrinkage to calibrate student trust in teacher logits. Operating on Qwen2.5-7B, RSIPO achieves 98.6% success on ALFWorld, outperforming standard GRPO by 17.4 percentage points.

Zoom

Recent attempts to scale tool-use post-training for autonomous agents have focused primarily on synthesizing standalone execution environments. However, environments constitute merely one component of a holistic agentic interaction system comprising environments, tasks, harnesses, and evaluators; scaling environments in isolation yields inconsistent training signals. Researchers from Fudan University introduce WEFT (arXiv:2609.36887), a whole-system evolution framework for tool-use post-training. WEFT coordinates environment breadth, task complexity, and interaction diversity, applying execution-driven self-evolution to attribute failures and iteratively revise flawed components. To ensure optimization stability and concurrency reliability, WEFT introduces prefix-preserving sampling, atomic-turn credit assignment, and MegaMCP—a Model Context Protocol infrastructure providing isolated, recoverable state management across concurrent rollouts. WEFT-8B and WEFT-14B outperform matched environment-scaling baselines across BFCL V4, tau^2-Bench, and Claw-Eval, with WEFT-14B beating Agent-World-14B by 6.41, 2.23, and 12.27 percentage points, while WEFT-35B-A3B excels on long-horizon benchmarks like Toolathlon-Verified.

Zoom

As autonomous LLM agents are entrusted with consequential real-world user decisions—such as selecting software packages, citing research literature, and booking travel accommodations—the neutrality and reliability of their autonomous choices remain an unverified assumption. Researchers from Seoul National University and NAVER AI Lab present the first comprehensive audit on 'Source Preference in the Wild' (arXiv:2610.03195). Evaluating 12 leading agent models across three practical domains with strictly controlled positional and feature parity, the study unveils a pervasive and statistically uniform source bias. Alarmingly, this brand/source preference frequently overrides strict user constraints: when an option meeting one fewer requirement originates from a favored source while the fully compliant alternative comes from a disfavored source, agents choose the objectively inferior option approximately two-thirds (~66.7%) of the time. The authors demonstrate that source tags trigger cognitive shortcuts when specifications are incomplete, and prove that providing structured information and injecting anti-preconception system prompts effectively suppresses this bias.

Zoom

Terminal agents execute stochastic model outputs directly in real bash environments, where an erroneous command (e.g., misconfigured dependency install, file deletion) irreversibly corrupts system state and derails downstream progress. Researchers introduce Mid-Harness, a paradigm that allocates test-time compute precisely at the boundary between model generation and the evaluation harness. By sampling and verifying candidate actions before sending them to the terminal—rather than expensively generating multiple full rollout trajectories—Mid-Harness elevates Pass@1 on TerminalBench-Lite from 50.00% to 68.03% under action scaling, achieving vastly higher reliability at substantially lower token cost.

Zoom

Alibaba Tongyi Lab (Tongyi-MAI, @Alibaba_Qwen) has open-sourced Qwen-Planner-Agent (arXiv: 2609.29892, Project: tongyi-mai.github.io/Qwen-Planner-Agent/), a next-generation mobile planner agent operating on Android systems. Overcoming the bottleneck of costly real-device interaction traces and brittle multi-turn reliability, Qwen-Planner-Agent establishes a closed-loop 'AI-for-AI' co-evolution framework. Integrating an agentic data flywheel, Competence-Aware Reward-and-Advantage Engineering (CARE) reinforcement learning, and execution-evidence model adaptation, it achieves top rank on MobilePA-Bench.

Zoom

StepFun (@StepFun_ai) released Step Code v0.1.0 as MIT open source. The company says one CLI covers the full loop from reading and editing code to running tests and shipping, with fast execution, high token efficiency, and long-horizon reliability. Reported scores: about 80.9% on Terminal-Bench 2.1 and 73.3% on their 150-task Multi-Frame benchmark, plus one-command static publishing via StepPage. Repo: https://github.com/stepfun-ai/Step-Code

Zoom