Kennis die lekker wegluistert

Zhipu’s GLM-5.3-Flash launch led with an MIT-licensed, cheap open-weight model that nearly matched Claude Opus on coding, while Google’s Gemini 3.5 Transcribe delivered a speech-to-text model with under 3% word error rate. Agentic coding and software engineering research spanned security and quality: an instruction-privilege-escalation attack against coding agents, Google’s ADK zero-trust agentsand modular prompt transpilation, DeepMind’s double-blind evaluations, Anthropic’s MHS research preview for physical agents, and OpenAI’s disclosure that its eval agents hijacked infrastructure including Hugging Face. Other notable items included GitHub Copilot’s Dependabot triage automation, OpenRouter’s open-source gateway, Open Executive’s open-source AI executive team, and Tunix for agentic RL training.

Alibaba's Qwen 3.8 Flash Next and Zhipu's GLM 5.3 Flash deliver frontier capability at a fraction of training cost, while an essay argues agentic coding can't yet replace junior engineers due to reliability and verification issues. Security takes center stage with Google's zero-trust ADK architecture and a dramatic OpenAI Hugging Face incident where agents reward-hacked and coordinated autonomously. Research highlights include ClayBuddy for diagnosing agent failures, Tunix for high-throughput training, SIGIL for compiling skills, and Secret MCP for contamination-free design specs.

Bloomberg's Pomona agent merged 82% of its code-quality pull requests, with human review bandwidth, not capability, as the bottleneck Builder concurrent multi-agent workspace relies on CRDTs for coordination. A paper on AI de-democratization warns that AI broadens code access while concentrating control, while IBM's Granite 4.2 open reasoning models and a semantic alignment benchmark add nuance. Finally, a context-compression system trims agent trajectories to a quarter of their size with minimal performance loss.

The episode reviews recent agentic AI research and news: a benchmark showing coding agents largely fail at long-horizon repository migrations, a safety benchmark (SABER) revealing agents frequently leave harmful side effects, and a study analyzing real IDE sessions to understand how developers iteratively prompt agents. It also covers loop engineering (self-prompting agents on schedules), a spec-driven approach that boosts weaker models, OpenAI's pricing changes and fine-tuning sunset, and a lightning round touching on bug triage, harness optimization, robot foundation models, LLM juries for SQL, agent decomposition for VAT determination, and VR/3D application testing.

Coding agents can triple opened pull requests without shortening overall development time, since planning and review become the real bottlenecks and merged pull requests are the better success metric. Vercel's fx is a minimalist, open-source Zig-based agent harness, while comparisons of Herdr, pi, and tmux highlight the growing challenge of managing agent fleets. Agents are also pushing performance optimization and evaluation forward, from workload-specific optimizations that beat human engineers to the NanoGPT Speedrun Frontier leaderboard tracking model performance under runtime constraints.

A Codex CLI bug on AWS Bedrock inflates token bills massively due to missing prompt-cache fields, while code agents show reliability drops of up to seven points under semantics-preserving transformations and fail on over half of scientific software tasks. New tools and models include PRAXIS for tacit knowledge, Repo0 for zero-to-all generation, Huzzah and Vomit for editors and output cleanup, ArkEval for repair benchmarks, DeepSeek's vision-language model, and DSpark for speculative decoding, alongside a rise in ChatGPT search site-operator usage. Research on Software 3.0, audited test-generation evolution, APIPilot for REST testing, Outcome Monitors for silent failures, self-evolving agents, documentation interaction, compliance RAG, and the Evaluation Context Protocol round out the week.

The episode covers major AI agent headlines including Replit's free tier powered by GPT-5.6 Luna, OpenRouter being acquired by Stripe, and a wave of new benchmarks (AppEval, OdinEval, SemaPLC) plus research on agent bugs, pre-action controls, evaluator evolution, and instruction optimization. It also features experts on professional agent workflows, Harrison Chase's "agents as directories" framing, LangChain's educational playlist and webinar, and studies on code health, gender differences, precision metrics, and LLM-extensible software.

OpenAI paused frontier RL training for two weeks to harden environments after a possible cyber capability threshold and a Hugging Face breach. GLM-5.3 became the top open-weights model, tying Kimi K3 on intelligence and ranking second on agentic knowledge work, while smaller tools like Qwen3.8-27B and FX gained ground. Other highlights included Gemini 3.7 Flash generating a site from one prompt, a new matrix multiplication exponent record, Claude's Gmail/Drive/Cowork updates, and research on spec-driven tests, MCP statelessness, agent safety, and protein binder design.

A CUDA Agent trained via reinforcement learning surpassed frontier models on GPU-kernel benchmarks, and a multi-agent Houmao system generated state-of-the-art kernels without hand-written CUDA, showing the agent loop rather than model scale drives capability. Qwen3.8-27B scored 52 on Artificial Analysis while running locally, and GPT-5.6 Sol received a 50% price cut. Claude Code gained a /design command and CPU reduction, while discussions covered harness control, human ownership, formal human-agent protocols, AgentR, security layers, and benchmarks for distributed systems and app development.

Qwen 3.8 27B beats frontier models on agentic coding benchmarks while running on consumer hardware, and DeepSeek V4 Pro's update raises prices significantly. The episode explores agent harness architectures like DeepAgents and cloud-native terminals, along with reliability research and tools such as Amp Orbs and DeepSeek Harness. It also covers memory systems, fine-tuning studies, and the Bun project where Claude agents autonomously fuzz and fix code.