Kennis die lekker wegluistert

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.

OpenAI released GPT-6 Astra, a hosted-only model with a million-token context and state persistence for computer use, alongside Meta’s Muse Image model and Anthropic’s open-source shopping agent blueprint. Google drove infrastructure updates including stateless MCP, session-aware load balancing, prompt transpilation, and agent platform evals, while studies covered tool selection, credit assignment, and evaluation gaps. The podcast closed with a lightning round on ARTEMIS, Tunix, DRACO, ARC-AGI-3 results, and SWE-Gate findings.

Quick gut check before we start. If I told you an AI just outscored every human at the world's top programming olympiad, would you even blink? Me neither, and that's basically this whole week. Welcome to The Daily Agentic AI Podcast, your daily look at agentic AI, autonomous agents, and the models pouring out of Google, Meta, Anthropic, and OpenAI. It's Thursday, September third, twenty twenty-six, I'm /jɑrnoː/, and Stephan is with me as always. This podcast is brought to you by Source Labs, the...

Anthropic released Claude Fable 5.1 and Mythos 5.1 as one model with different safeguard layers, doubling agentic science benchmark scores but increasing per-task cost through heavier token output, alongside satirical release notes from Simon Willison. OpenAI's Astra became the first model to hit the Critical cyber-risk tier under its Preparedness Framework after demonstrating autonomous vulnerability discovery and exploit chaining. Other coverage included coding-harness engineering, small critic models for code agents, an autonomous Blender demo with GLM-5.3-Flash, Paint.NET's Claude-written Direct2D rewrite, spec-driven development, zero-trust agents with ADK, Perplexity's hybrid Mac compute, Genkit Go skills, persistent runtime-independent agents, Tunix, HarnessDev, Qwen3.8-Max, Gemini Enterprise Agent Platform GA, PTA-IRT, and SWE-bench Science.

The podcast covers agent self-improvement with Ouroboros, a verification tool study with limited benefits, and a misalignment analysis of coding agents. It also discusses Agent Plugins adoption, zero-trust security, Google's agent evaluations, legacy migration, RL training with Tunix, reward hacking in Hacker-Opus, and automated repair agents.

The episode opens with a researcher demonstrating a full prompt injection attack chain against Claude Code's Auto Mode, contradicting Anthropic's zero-percent security claim, then pivots to OpenAI's ChatGPT Work agentic tool and a study showing Claude Code plugins are largely co-written by Claude itself. It covers the architectural convergence of GLM 5.3 Flash and Qwen 3.8 Flash Next, Gemini Omni 1.1 Flash's video editing, EvoRepair's self-evolving vulnerability patching, SpecMine and Conductor for spec-driven development, Genkit Go's progressive-disclosure Agent Skills, LangChain's stateless MCP support, and Apple Silicon performance gains. It closes with the Model Hardware Standard, set-shifting behavioral tests, EnvHarness adaptive training, agent civilization experiment critiques, and grim findings that mean-time-to-exploit has gone negative.

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.