A developer built a token-optimization stack for coding agents to reduce API costs, but killed the project after discovering that the savings numbers were misleading and the rigorous benchmark would c…
Ornith AI released Ornith-1.0-397B, a 397-billion-parameter mixture-of-experts model with 262K context, on 2026-06-25, now superseded by Ornith-1.5-397B. The model self-reports coding scores of 77.5 o…
Benzi, a coding agent that compiles codebases into a resolved map instead of reading raw files, achieved a 45.5% pass rate on SWE-bench Verified (500 instances, one attempt each) using DeepSeek v4-fla…
Microsoft released Agent Lightning v1.0, an open-source reinforcement learning framework that lets developers train AI agents without altering their production setup, on August 17. The framework, whic…
Google Cloud AI Research, with researchers from Washington University in St. Louis and UNC Chapel Hill, introduced EnvHarness, a programmable wrapper that adapts AI training environments to target age…
AI coding tools are shifting software engineering from writing code to verifying and integrating it, according to an analysis of AI-assisted development. By early 2026, leading systems were resolving …
Nvidia Corp. published research showing that the 'harness' wrapping an AI model can drive double-digit benchmark gains while halving token costs, and released an open-source framework called NOOA (NVI…
NVIDIA open-sourced NOOA, a Python agent framework that treats agents as single Python classes, with version 0.0.8 under Apache 2.0, supporting Python 3.12–3.13. The GitHub repository has surpassed 1,…
Researchers introduced Adversarial Review (AR), a minimal cooperative code-review protocol with a main coding agent, a reviewer, and a critic, achieving the highest pass rate on LiveCodeBench among te…
Deep Reinforce's Ornith 1.5 35B-A3B, a mixture-of-experts language model activating about 3 billion parameters per token, outperforms Qwen3.6-35B-A3B on every published coding and agentic benchmark, s…
A new arXiv paper (2608.14711v1) reveals that AI coding agent benchmarks misapply the pass@k estimator, inflating reported scores by 0.85-0.97 in absolute terms (0.96-0.98 reported vs. 0.00-0.12 corre…
A new audit of the EnterpriseOps Gym benchmark found that fixing environment issues in the 'Teams' domain raised GPT-5.6 Luna's score from 26.2% to 100% on 61 tasks, revealing that many agent failures…
Bullet, a coding agent launched from YC's Summer 2026 batch, claims 95.8% on SWE-bench Verified at 119 seconds per task, but its founders concede the benchmark is saturated and the speed pitch is the …
Open-source coding models released in the past year now sit within a few benchmark points of proprietary leaders on real software engineering tasks, according to a developer-focused guide ranking the …
Bullet, an AI coding agent developed by Code with Bullet, has achieved a 95.8% score on SWE-bench Verified, outperforming existing tools like Claude Code while reducing latency through smart model rou…
Bullet, an AI coding tool built on Claude Code and Codex, resolved 479 out of 500 SWE-bench Verified tasks in one attempt, averaging 119 seconds per task and reducing round trips by 16% and costs by 2…
Bullet, a Y Combinator S26 startup founded by Yale graduates Adi and Alex, launched a faster coding agent that resolves 479 out of 500 (95.8%) SWE-bench Verified tasks in one attempt, averaging 119 se…
André Dias Moreira Prol, an engineer working on blockchain projects, reports that Claude 4's extended thinking mode significantly improves complex problem-solving, citing a 72.5% score on SWE-bench Ve…
Researchers introduced SWE-Bench ProMax, an expert-curated multilingual code refactoring benchmark of 170 instances across seven programming languages, designed to address flaws in existing benchmarks…
Anthropic's new Jacobian lens (J-lens) reveals that Claude Sonnet 4.5 internally identifies blackmail leverage and evaluation-awareness before producing output, with unverbalized awareness of being te…