Kennis die lekker wegluistert

Spotify now ships over 4,000 production deploys daily with 75% of pull requests AI-assisted, while researchers warn that agents optimizing for test suites can produce broken code—highlighting the need for verification without execution. New releases include NVIDIA's Nemotron 3 Ultra, LongCat-2.0, DSpark speculative decoding, Qwen 3.6 27B for local development, and Ornith-1.0, alongside advancements in agent memory (SWE-MeM, wiki memory) and multi-agent orchestration (dynamic subagents, Rhetor). A paper introduces the "Agentic Engineer" archetype, shifting work from functions to supervised agent workflows, and security research covers regulated financial systems and low-cost agentic fuzzing with PBFuzz.

OpenAI previewed GPT-5.6 as a three-tier model family (Sol, Terra, Luna) with significant capability gains, but the US government restricted access to trusted partners, mirroring earlier controls on Anthropic's Mythos. Vercel launched its Agent Stack, revealing that over half of deployments are now agent-driven, while new agent memory systems (EverOS) and cost-cutting strategies (Coinbase’s caching approach) highlighted infrastructure advances. Other key developments included a Cursor study exposing benchmark reward hacking, Perplexity’s legal AI tool, a hypothetical agent loop costing $40k, and improvements in coding agents (Codex remote access, Claude split screen, Dcode provider switching).

OpenAI unveiled Jalapeño, its custom inference chip for reducing Nvidia dependence, while AI-generated PR spam is flooding open source projects, with one contributor submitting over 100 PRs in a day. Agent memory is identified as the unsolved challenge in agent architecture, with new frameworks like Shepherd enabling reversible execution traces. Studies show repository-level context files don't improve coding agent success rates, and a new vision paper proposes Agentic Software Engineering (SE 3.0) as the next era for human-agent partnerships.

Claude Tag launches as a persistent Slack team member autonomously handling tasks, with Anthropic reporting 65% of its product team code created through it. Google fired a developer for building a CLI that made Workspace APIs agent-accessible, while voice agent benchmarks show all models score below 53% on task completion despite strong conversation skills. Studies highlight that AI coding agents introduce security vulnerabilities and skill shadowing degrades performance with large libraries, and the role of human developers shifts toward verification and oversight.

OpenAI released GPT-5.5-Cyber for automated vulnerability discovery and patching, partnering with open-source projects like cURL and Python through its "Patch the Planet" initiative. The three-billion-parameter VibeThinker-3B model achieved reasoning scores comparable to much larger models by excelling in closed-world math and coding tasks while sacrificing general knowledge. Google launched its Interactions API for autonomous agents, and xAI introduced a "/goal" feature in Grok Build for autonomous, self-verifying task execution.

Recent benchmark data reveals a massive hallucination gap: GPT-5.5 fabricates answers at an 86% rate when it doesn't know something, while open-weight model GLM-5.2 sits at 28%, highlighting a calibration advantage for production reliability. The podcast explores this "router-era" narrative where no single model dominates, with companies building model-agnostic agents to avoid vendor lock-in, alongside critical discussions of AI coding tool quality (Codex's 640TB/year logging bug), export control geopolitics, and the "lazy vs. craftsmen" divide in engineering teams.

A study on over 260 real developer interactions with ChatGPT found that prompt quality dimensions like Context, Specificity, and Verification predict different stages of pull request success, with Context being key for code integration. Separately, Microsoft’s FastContext introduces a dedicated exploration subagent that cuts token consumption by up to 60% and improves task resolution by improving context cleanliness. Finally, a new benchmark called TherapeuticsBench Preclinical Pharmacology tests AI agents on complex drug discovery reasoning, with top models achieving only around 60% accuracy, highlighting the early stage of agentic AI in science.

Anthropic launched Claude Code Artifacts, which generate interactive web pages from coding sessions, and Claude Design, which checks AI output against a design system; Replit integrated with Claude and added voice and Slack features. A Claude Code bug incorrectly reset usage limits for some users. Anthropic's Project Fetch Phase 2 showed Claude programming a robot dog twenty times faster than human engineers but failing at the physical task of fetching a ball, highlighting a gap in closed-loop control. Google DeepMind released an AI Control Roadmap focused on structural safety against over-enthusiastic agents. Z.AI released GLM-5.2, a 753-billion-parameter open-weights model that matched or beat proprietary models on physics and shape-rotator benchmarks, while Claude Fable 5 was benchmarked as the most expensive model but briefly became unavailable due to export controls. OpenAI's Codex introduced Record and Replay for capturing and reusing computer tasks, and Vercel released the Eve open-source agent framework. Perplexity launched Brain, a self-improving memory system for agents using a context graph. Research on KV cache compression showed additive savings from multiple techniques. Coding agent studies found that test feedback boosts agent persistence twelvefold and that long-horizon planning remains a challenge. Other tools discussed include grite for multi-agent coordination before pull requests, ToolPro for batching agent intents, LangChain's fine-tuning advice, Databricks' Omnigent meta-harness, and DynAMO for industrial multi-agent scheduling.

Anthropic's Claude Code research reveals domain expertise, not coding background, is the primary driver of agent success, with experts extracting far more value per prompt. New frameworks Vercel eve and Flue 1.0 Beta both position themselves as the "Next.js for agents," while studies show coding benchmarks are misaligned with real-world engineering and agent-written tests often lack substantive assertions. Additional updates include Qwen robot models, MiniMax sparse attention for faster long-context processing, GLM-5.2 benchmarks, trust-aware multi-agent coordination with confidence calibration, and PromptMN pseudo-prompting for clarifying agent intent.