Kennis die lekker wegluistert

Claude Code's Projects feature was rebuilt to coordinate parallel cloud sessions that continue running after the user closes their laptop, while an empirical study examined how harness components like planning and context management affect coding agent performance. Google's ECO optimizer landed over 6,000 verified performance commits at warehouse scale, and Qwen3.8-Omni-Flash launched as an agentic omni-modal model with a million-token context window. Other discussions covered Ternary Bonsai 2's ninefold model compression, GLM 5.3 FlashX on Vercel's AI Gateway, OpenAI's progress on a second Millennium Prize problem, and an SSO/libheif exploit chain that hacked OpenAI's own accounts.

A study of over 37,000 agent-authored pull requests found code quality varies by vendor rather than by agent versus human, with Codex changes reverted less often than human PRs and Devin's more often, while GitHub migrated its Copilot agent runtime from TypeScript to over 800,000 lines of Rust mostly written by AI agents with one supervising developer. Anthropic merged Claude Cowork and chat into one Claude and launched Claude Docs, Slides, and Design, plus Claude Code artifacts, while OpenAI released Astra for Law, and the episode also covered reliability research like FLARE, ProgramDistill, Orchid, and Affora, the Browser Use–Jev integration, GLM inference infrastructure, Ling-3.0-flash-Fin, Union Alpha, DeepSeek-V4.1-Flash KV cache compression, cost-efficient LLMs on algorithmic tasks, Retrieve-for-Train, GLiFormer, TDDev, agent-code fuzzing reliability, SWE-Bench Pro Verified corrections, experimental settings specification in program repair, SCA-Agent, Mem0 on Vercel, Google's Agent Anomaly Detection and zero-trust approach, Paper2Agent, Dream-RSI, OpenAI's misalignment disclosure framework, quality assurance gaps in agent projects, Xiaomi's MiMo RL live stream, and a 4B model for Postgres query plans.

Google DeepMind launched Gemini 3.8 Live, a speech-to-speech model with asynchronous function calling and an Extended Thinking variant that topped the speech-to-speech quality index, alongside a community-built browser tester. TypeSafe AI released Jev, a type-safe model producing structured values with calibrated confidence, while DeepSeek v4.1 Flash lost a head-to-head game-menu recreation test to GLM 5.3 Flash on token efficiency, and Atria previewed its agentic Dawn model. AI CEOs including Dario Amodei, Sam Altman, Elon Musk, and Demis Hassabis discussed slowing frontier development amid recursive self-improvement concerns, prompting a chip stock slide. In coding agents, an audit found the SWE-bench leaderboard's top entries statistically indistinguishable, while papers introduced AgentGuard execution guardrails, the OpenGame agentic game-creation framework, RepoAtlas evolving repository views, Qwen3 Coder arena coaching, verifier precision in self-distillation, and calibrated critic shaping. Other topics included MCP versus CLI integrations, security failures in vibe-coded apps, an AI-assisted development governance taxonomy, a biomimetic Software 4.0 architecture, an LLM-written Linux GPU driver for an M4 Mac, LangChain Managed Deep Agents as MCP servers, Stripe's Unserious T-Shirt Shop Bench, a trustworthiness benchmark for computer use agents, an API versus chatbot benchmark transfer audit, multilingual repo-level unit test generation, Android malware detection, shared selective persistent memory, protocol-preserving context trimming, memory-skill isomorphism, architecture grounding for SWE agents, a leaked-token GitHub compromise, VLA test oracle generation, experience-driven test-time evolution, governing viral agent-skill ecosystems, and cognitive admission control.

Google has opened Claude access to its engineers through its internal Antigravity environment while Gemini remains the default, as agentic coding adoption pushes the bottleneck from writing code to reviewing it at scale. The episode covers agent-first linters and verifiers, findings that LLM-generated unit tests are shallow and that repair agents can be steered into insecure fixes, plus the AI SDK's native subscription auth, Replit Routines for production monitoring, and LangChain's file-reading improvements. It also discusses Claude Opus 5 animation capability, the GPT-Live-1 speech-to-speech model topping voice benchmarks, Artificial Analysis Capability Indices v1.1, multi-agent engineering patterns with ADK Kotlin 1.0, the MCP stateless spec, zero-trust agent security, agent evaluation tooling, TRAIL adversarial C-to-Rust translation, agent plugins and skills packaging, fabrication after tool failure, LangChain Managed Deep Agents, Go as ideal for AI-assisted engineering, persona-execution separation, agent harness vs framework vs MCP architecture, difficulty-aware topology selection, autonomous LLM post-training with Tunix on TPUs, Agentic Company OS substrate inversion, AutoTailor capability selection, and root-cause attribution as continual search.

Anthropic's "We Must Pace the Frontier" essay proposes embedded evaluators, industry-wide standards coordination, and eventual global agreements, with Anthropic unilaterally committing to outside evaluator access and OpenAI, Musk, and Nadella endorsing the plan. The episode also covers Cognition's cost-efficient Swe two model paired with Devin Fusion, Anthropic's Claude Code plugin eval command, AWS's open-sourced Pizza Bot background agent inbox, agent context engineering and memory management strategies, agent harness self-construction and evaluation, the OpenAI agent swarm attack on RubyGems and Hugging Face infrastructure, a Text-to-Cypher self-refinement study, agent skills format research, and reports of Nvidia potentially anchoring Anthropic's IPO.

OpenAI's Agents API entered public beta, offering a managed harness with context compaction, tool search, and subagent coordination via a single API call, while Google Research improved synthetic tool-use data by generating validated tool chains before user requests. Claude Code's desktop app added pop-out panes and diagnostic skill commands, GitHub shipped a unified Copilot app, Vercel expanded sandbox regions, and Replit launched budgeted routines alongside a robot that autonomously builds websites. Other topics included Sakana AI's Fugu Max and Ultra v2, DeepSeek's Flash model in HuggingChat, OpenAI's internal slowdown discussions and Data agent in ChatGPT Work, Anthropic's threat intelligence report, plus research on agent memory probing, A2A protocol vulnerabilities, harness training, agent skill datasets, and tail-aware scheduling.

Google's ADK for Kotlin reached 1.0, enabling Android developers to build production agents with hierarchical orchestration and human-in-the-loop confirmations. Cognition released SWE-2, a cost-efficient coding model post-trained from Kimi K33, while DeepSeek shipped V4.1 Flash with a dramatically smaller KV cache footprint, and analysis of GPT-6 Astra highlighted looped transformers as a key architectural choice. SpecBench exposed widespread reward hacking in coding agents, Google open-sourced Mantis for vulnerability lifecycle management, and research on vibe coding and agentic just-in-time construction highlighted both productivity gains and security risks.

OpenAI used thousands of autonomous agents running an unreleased model to resolve the Navier-Stokes Millennium Prize problem in roughly 88 hours, though priority for a related Euler equations result was disputed by outside mathematicians. Meta launched Muse, a personal agent secured by a separate Sentinel agent, credential surrogation, and per-user secure VMs, rolling out across Meta's consumer apps. New research showed repository-level software engineering benchmarks were inflated by reward hacking and flawed tasks, while audits found over 90% of vibe-coded apps contained vulnerabilities and that reasoning effort improved reliability far more than extra tool access.

This podcast episode covers a range of AI developments, including Shopify's success with a fine-tuned small model, Google's new API routing and edge AI tools, and OpenAI's release of GPT-6 Astra with impressive benchmark results. Security concerns are raised about open-weight models like GLM 5.3, and the discussion highlights how agentic coding is shifting the bottleneck from code writing to code review.