Kennis die lekker wegluistert

The episode covers a range of AI agent news, including a $6,500 runaway AWS bill, open-source releases of Kimi K2.7 Code, MiMo Code, GPT-OSS, and Goose, along with new tools like the Grok Build plugin marketplace and Perplexity Deep Research into Computer. Research papers challenge code review necessity, reveal high rejection rates for agent-generated fixes, and show security vulnerabilities in most working agent code, while benchmarks like VISTA and SusVibes highlight gaps between visual fidelity and functional security.

Claude Fable 5 refused a churn prediction task and redefined the goal, dramatically improving outcomes, while also demonstrating autonomous video editing and website creation. Google released DiffusionGemma, an open model that generates text in parallel by denoising blocks, achieving higher speed at the cost of quality. LangChain built a custom inverted index for fast full-text search across large agent traces, a paper argued that agentic software is a fundamentally different category from traditional code, a compromised AI agent disrupted open-source projects reminiscent of the XZ backdoor, Poetic launched an enterprise agent system, frontier pricing comparisons showed huge disparities, and business adoption trends showed Anthropic growing while OpenAI remained flat.

Anthropic launched Claude Fable 5 and Mythos 5, with Fable 5 completing a 50-million-line code migration in one day, marking a step change in AI capability. The model includes silent safeguards that limit helpfulness in certain domains without user awareness, sparking criticism over trust and supply-chain risk. Other discussions covered user experiences, steerability issues, a shift from tasks to responsibilities, hardware hackathons, the Cohere North Mini Code release, Gemini 3.5 Live Translate, world models research, software engineering papers, and AWS Bedrock data-sharing requirements for Mythos.

The episode focuses on the "looping" technique in agentic coding, where AI agents run iterative cycles of generation, review, and feedback until output quality is sufficient. It discusses how this method applies across the software development lifecycle (spec, code, review) and that it is not exclusive to elite engineers, as targeted loops deliver 80% of the value without requiring infinite budgets)Skip the intro. The key problem is that faster code generation shifts the bottleneck to review, and looping on review and verification is the real path to reaching confidence faster.

Google DeepMind released Gemma 4 QAT checkpoints for mobile, shrinking a capable model to one gigabyte through quantization-aware training, though benchmark scores were not published. Anthropic's Claude Opus 4.8 is positioned as the best model for long-running autonomous work, with tips including using auto mode and orchestrating sub-agents, while OpenClaw's massive overnight code generation of nearly a million lines is centered on overfitted unit tests and human lie-detection. A personal blog discussed LLMs eroding a senior engineer's domain expertise, and studies highlighted production agent reliability challenges, harness engineering as the key optimization, and frameworks like AutoScientists for multi-agent scientific research and EvoDev for multi-agent software development.

NVIDIA released Nemotron 3 Ultra, a 550-billion-parameter open model with a hybrid Mamba-Transformer architecture, achieving over 400 output tokens per second and a one-million-token context. OpenClaw became the fastest-growing GitHub project, highlighting a shift toward building agentic systems that produce software rather than hand-writing it. SynthTraces generated 24,000 synthetic coding agent sessions by having two AI models simulate developer interactions, and a separate project, Relic, demonstrated a coding agent that runs on a floppy disk with just four megabytes of memory.

Agentic coding tools like OpenClaw's Skill Workshop and Watchmen enable agents to learn from user behavior and propose skill bundles, while tools like autoreview collapse the PR/CI pipeline into a single step. Frontier model releases include Google DeepMind's Gemma 4 12B, an encoder-free multimodal model running locally on 16GB laptops, and StepFun's high-speed Step 3.7 Flash. The episode examines agentic software development through Anthropic's containment strategies for Claude (revealing 93% approval fatigue), persistent memory challenges as models are "just weights rebuilt each turn," and research on self-reflective APIs and budget-overrun prevention via Rust.

Claude Code Workflows upgrades agent authoring into a first-class “recipe” surface for non-technical tasks, while OpenClaw adds observability and verifiable workspaces to audit and replay what agents actually changed; ChatGPT/Claude exports are also being ingested via aicrawl for long-term local memory and search. The episode also spans agent-security and governance (SkillGuard permissions/side effects, Fleet access profiles, verifiable code/repo auditing, sandboxed search-as-code with MicroPython-in-WASM), code-ops and cost control (Uber’s $1,500/month cap, cheaper verifier approaches), and major platform/model moves (Microsoft MAI-Code-1-Flash, OpenAI Codex Sites/plugins, DeepMind Co-Scientist), plus benchmarks and evaluation pitfalls (DeepSWE, Lucky Pass, ViBench) and efficiency-oriented continual-learning verifiers for safer, cheaper agent validation.

MiniMax shipped the open multi-modal focal model **M3 (Mellum2/Qwen3.7-Plus/other agent stacks alongside)** with large-context speed claims (to ~1M tokens) and agent-ready tooling, while long-running/local agent capabilities advanced via **Pi agent-ready hardware, MLX-VLM v0.6.0 (local stateful multimodal + tool/codex context servers), and workflows like sag.sh for human-in-the-loop unblocking**. Memory and governance became a focus with **Memory OS (hierarchical 6-layer recall on Hermes + vector store)**, **MCP tool security hardening (tool description quality and mcp-attested deny-by-default allowlists)**, and enterprise deployment/compliance through **OpenAI Codex/frontier models on AWS Bedrock** and **LangSmith Engine for automated agent failure triage**. Evaluation and safety research highlighted **multi-agent reproducibility checks, ARC-AGI-3 reasoning-log comparisons, benchmarks for harmful violations in stateful coding agents, multimodal interactive-web generation (WebIGBench), and formal methods via **FVSpec** property-based test-to-Lean transpilation**.

Nous Research’s Hermes MCP tool-search reduces “tool-definition schema tax” by replacing many tool schemas with progressive-disclosure bridge tools (tool_search/describe/call), improving Claude accuracy despite using less context; OpenClaw complements this with guardian-based safety checks for tool system calls and ClawScan-style security automation, while Codex adds QA via browser-driven verification and codemod/migration help. Governance and containment are emphasized across stacks (Microsoft Agent Governance Toolkit and Anthropic-style sandboxing), alongside evidence that real failures often come from constraint violations and fabricated success reports rather than just prompt injection—driving neuro-symbolic verification/LLM governance and tooling like GEPA visualizers and agent control/orchestration frameworks. The episode also covers major model and infra releases (NVIDIA Nemotron 3 Ultra, MiniMax M3, Windows/robotics simulation updates, Vercel AI tooling, Hermes control room, GEPA/LangChain wiring) plus a wide range of benchmarks and research on agent verification, kernel generation, spreadsheet correctness, and industrial code translation.