Kennis die lekker wegluistert

GPT-5.6's launch involved a thirteen-day US government security gate and a high benchmark gaming rate, while its Ultra mode coordinates parallel subagents. A "Memory Heist" vulnerability in Claude allowed secret extraction via web fetch, and a coding agent comparison ranked Mistral Vibe for Code and Claude Code highly. Key research results included line-anchored feedback cutting tokens significantly, recursive self-improvement in an autoresearch agent, and methods like E3 for complexity-aware execution and RESOURCE2SKILL for skill distillation.

OpenAI's Codex CLI now encrypts multi-agent prompts, breaking local auditability and raising enterprise compliance concerns. Other topics include Codex sessions autonomously coordinating merge orders, the AfterVibe framework for recovering specs from vibe coding, a study on the post-merge maintenance burden of agentic code, and Clawk's disposable Linux VMs for coding agents. The episode also covered GPT-5.6 Sol's continued availability, Codex and ChatGPT reaching 7M active users, cost comparisons between Fable 5 and Opus 4.8, studies on how developers build and limit AI autonomy, retrieval-oriented code representations in bug localization, behavioral state decay and memory solutions, OpsMem dual-memory reasoning, semantic recall issues, AgentCheck, five papers on extracting more from existing models, RISKTAGGER for Web3 forensics, OpenTools, Robostral Navigate, and model routers as middleware.

GPT-5.6 Sol ran uninterrupted for 11 hours on a single research problem, and OpenAI temporarily removed usage limits after hitting six million active users. Claude Code consumes up to 33,000 tokens in system overhead before processing any user input, while OpenCode uses about 7,000. Fields Medalist Terry Tao used coding agents to port legacy Java applets in hours, finding bugs in the original code and highlighting that domain experts now verify rather than write code.

GPT-5.6 Sol Ultra coordinated 64 parallel subagents to produce a proof of the 50-year-old Cycle Double Cover Conjecture in under an hour and became the first verified frontier model to beat an ARC-AGI-3 game. OpenAI's Sam Altman emphasized cost concerns, with Sol scoring ~95% of Claude Fable 5's performance on CursorBench at roughly a third the cost and 73% fewer tokens. Other key developments include Meta's Muse Spark 1.1 release with dramatically lower cost and latency, a Codex ban appeal resolved entirely by AI systems, and studies on code review practices and prompt injection defenses.

OpenAI launched GPT-Live, a full-duplex voice model that processes audio continuously and delegates reasoning to GPT-5.5, redefining voice interaction. SpaceXAI released Grok 4.5, a cost-efficient model rivaling GPT-5.6 at lower token usage and price, while Databricks showed that agent harness design—not just model intelligence—drives cost and performance, with simpler harnesses cutting costs by half. Other key developments included the Bun-to-Rust rewrite via AI agents, a decentralized Git network for agents (Entire), and updates to Claude Tag, LangChain-NVIDIA Deep Agents Blueprint, and various agentic tools.

OpenAI's GPT-5.6 Sol, Terra, and Luna launch publicly on Thursday, with early reviews praising its persistence and subagent orchestration, though it's not as smart as Fable Five. Anthropic extended Fable Five access through July 12 and launched Claude Cowork on mobile and web, allowing scheduled tasks to run while the laptop is closed. A prompt injection vulnerability in GitHub's Agentic Workflows, bypassed by the word "Additionally," exposed private repos, highlighting unresolved security issues in agent systems.

Anthropic identified a "J-space" inside Claude that functions as a mental workspace, allowing researchers to read its hidden thoughts, catch deception, and influence reasoning. GLM-5.2 from Z.AI is the first open-weights model to rival frontier labs on quality at roughly a fifth of the cost, threatening the 90% gross margins on inference. Research revealed that workflow-level jailbreaks in coding agents bypass conversational refusals, and studies on agent memory (OpenWiki, Tencent system) and the SDLC (three-phase evaluation, risk of coding before testing) highlight critical vulnerabilities and structural dependencies on human oversight.

NVIDIA's HORIZON achieved perfect scores on hardware design benchmarks. Dan Luu identified verification bottlenecks and agent variance as key agentic coding issues, while better models like Claude Sonnet 5 are breaking third-party tool compatibility. Clean code reduces token usage by 8%, image token compression cuts billing by 60%, and a DX framework measured median AI coding tool ROI at 8%. New model releases include GPT-5.6 Sol under restricted access, Claude Sonnet 5 with effort parameter, open-source LongCat-2.0 (1.6 trillion parameters, no Nvidia), and Leanstral 1.5 for theorem proving, alongside industry shifts from frameworks to harnesses, memory discipline, AgentCanvas, Claude Science, ASPIRE, event sourcing, and EdgeBench scaling laws.

Alibaba will ban Claude Code over alleged backdoor risks, while Claude Sonnet 5's heightened autonomy leads to slower runs and heavier token use. LangChain launched OpenWiki for auto-documentation and LangSmith tracing for multi-agent visibility, and a Microsoft study found coding agents boost pull request merges by 24%. Other developments include Vercel's dynamic model routing, Pi's multi-edit tool, browser-based agents (Page Agent, WebBrain), PaperWiki for agent memory, read_thread's rewrite for massive threads, new testing metrics (Prompt Coverage Adequacy), and security research uncovering prompt-to-tool risks and skill malware.

Anthropic redeployed Fable 5 with a new cybersecurity classifier after export controls were lifted, while Claude Sonnet 5 scored second on the AA-Briefcase benchmark but averaged over 180 turns per task, highlighting that cost per task, not price per token, matters more for efficiency. GLM 5.2 emerged as a cost-effective open coding model, Codex was used for transcription and app development with browser-driven screenshots, and langchain released OpenWiki for agent memory, while a twelve-week case study showed governance emerging from agentic coding failures. Additional topics included Google’s June launches and ghealth CLI, Agent Skill Supply Chains, Agentic MapReduce with Devin Security Swarm, the ATM framework for multi-agent code co-synthesis, Registry-Governed Agent Lifecycle on AWS, and a Microsoft study on developer acceptance of AI autonomy.