Tag
A cross-model analysis tested how four open coding models (Qwen3.5-9B, GLM-4-9B, Gemma, and Nemotron-Nano-9B) complete a truncated curl upload URL, finding that cloud-provider bias and fake credentials appear mainly in Qwen's tail outputs while NVIDIA's model shows the best hygiene, with a public harness and 400,000-row dataset released.
Particle-1, a coding-agent-focused model emphasizing token efficiency (~45% fewer output tokens than DeepSeek V4.1 Flash with comparable accuracy), is now available in OpenClaw v2026.10.10, with zero data retention for standard inference.
A long-time AI user reports that models across ChatGPT, Grok, Claude, DeepSeek and coding agents like OpenCode have noticeably worsened at task understanding and instruction-following over the past year, prompting questions about post-training, alignment and inference changes. The post asks the community whether others are seeing the same cross-provider regression.
A one-week case study embeds Claude Code and Codex CLI as party members on an offline WoW 3.3.5a server, yielding design lessons like 'sight without authority', typed action vocabularies, explicit refusal signals, and schema design that follows natural speech.
An empirical arXiv study examines /retro, a skill that lets coding agents review completed sessions and propose harness changes, finding that while many proposals are generated, only some reduce repeated work on later tasks — showing harness self-evolution needs mechanisms to verify and revise changes.
Revolut moved ~2,000 engineers onto self-hosted GLM models for coding, retiring Claude Code and Codex to an exception list — running on 56 Nvidia B200s at a blended 30 cents per million tokens (about a tenth of prior Anthropic/OpenAI spend), with engineering lead time improving 70%.
A post argues that coding agents often claim to fix security bugs while leaving the original PoC still exploitable, and proposes freezing the PoC, evidence, and missed check as fixtures for every rejected action so patches are verified against the trigger, not just the diff.
Nana is a new terminal designed to run every coding agent on Windows, launched via Product Hunt.
A developer building an independent verification prototype for coding agents shares two reproduced cases where pytest reported passing despite failing tests (exit status 0, config-based deselection) and asks the community how they verify agent-generated changes before accepting them.
A developer rant arguing that tools like vercel-labs' skills installer and Docker Gordon ship unconfigurable hardcoded timeouts that break users in slower-network environments, and that every timeout should be safe, documented, and overridable.
A tip suggesting developers prompt their coding agent to compare original requests with the built implementation and surface up to three consequential user-facing decisions the agent made on its own, across tools like Claude Code, Codex, Cursor, Lovable, or Replit.
A blog post arguing that coding-agent session transcripts and trajectories are valuable records for debugging failures, understanding costs, and reusing fixes — and warning that Claude Code's 30-day cleanup can delete them permanently. It breaks down what a session archive contains and cites an RTK benchmark attempt that passed only after 339 consecutive errors at ~9x the baseline cost.
Momentic (YC W24) launched on Launch HN, showcasing Mo, an AI QA engineer that tests apps end-to-end and has caught over 117k bugs across 70M+ test runs. The platform integrates with coding agents via MCP and CI pipelines using plain-English YAML tests with self-healing capabilities.
Beacon, an open-source self-improving memory and context handoff tool for coding agents, is trending #1 on Product Hunt, letting developers resume sessions across 20+ tools like Claude Code, Cursor, and Codex without re-explaining the project.
MiniMax thanked community builders who shrank M3.1 Flash open weights from 252.4 GB to as low as 93 GB in 10 days, with quantized variants ranging from GGUF to EXL3 and MXFP4 delivering strong accuracy and throughput on consumer and datacenter hardware.
Google's developer blog argues for testing discrete agent behaviors instead of relying only on composite benchmark scores, framing eval suites mainly as regression guards against prompt, tool, or model changes; the post contrasts this with OpenAI's build-side harness engineering approach.
A curated list of tiny-to-medium models (mostly under 35B) suited for running LLM agents on a normal computer without a high-end GPU, ordered by capability and highlighting models with vision support.
The user recommends Factory’s technical article “Droid 1.0: Building a Frontier Harness,” highlighting engineering optimizations such as Programmatic Tool Calling, on-demand tool loading, long-running background tasks, and Agent Factory, and sharing a case in which the same optimization had opposite effects on costs for different models (Opus vs. Kimi).
An empirical study finds that continual harness self-evolution of coding agents across multiple software engineering tasks causes severe catastrophic forgetting, with SWE-Bench performance dropping from 94.0% to 32.0% after evolving on FeatureBench, and offers insights for future continual agent-harness evolution.
Qodo launched a Codex plugin that injects organizational codebase context and independent AI code review into developers' coding sessions, helping catch cross-repository P0 bugs before pull requests are opened.