Kennis die lekker wegluistert

Spawn is a multiplayer game world developed autonomously by a swarm of Fable-powered agents that spend over $3 million monthly on tokens, ship to production hourly, and roll back broken changes in about 30 seconds. A new paper finds coding agents overwhelmingly follow repository precedent over new task instructions, with file placement dramatically shifting adoption rates, and a trojanized Claude Code extension on Cursor's marketplace affected up to 5,000 users and attempted to install a persistent SSH backdoor. Other topics include Claude Code's subagent effort environment variable, Kino video editing with Claude Code and Codex, Meta's Muse Spark Open 816B model, Baseten's joey-1 decision model, Samaya's Sterling finance agent model, OpenAI API credits for subscribers, the Phonic 1 speech-to-speech model, Anthropic's 128 MB API request limit increase, DeepSeek V4.1 Flash local inference performance, Anthropic's server-side rule restricting game recreation, LangChain's Managed Deep Agents browser and async subagents, OpenAI's expanded frontier model monitoring, multi-agent swarms in Minecraft and Mindustry, SWExploit red-teaming, oversight-loosening directives in agent skills, the CARI runtime injection attack, and the Artificial Analysis Search Index.

Google's Gemini 4 Carbon launch claims frontier long-horizon agentic and cyber-defense capability with top benchmark scores, but it is initially restricted to vetted cyber defenders via the Fairwind Program, and Artificial Analysis notes a hallucination rate of roughly 11% versus Argon's 4% with mostly self-run, unreplicated figures. Mistral released open weights for the trillion-parameter multimodal Large 4 (Le Chonk) under modified MIT with a commercial license required for model-as-a-service, while GitHub Copilot Fusion orchestrates Anthropic and OpenAI models to reach within two points of Opus 5.5 on an agentic coding index at about half the cost and ~40% less time per task, though it trails on some benchmarks. Other discussions covered Vercel CI beta with agent-controllable runs, GitHub's Octoverse 2026 report (nearly 18 billion pushes, over 40% of PRs agent-started by September), MiniMax's open-source MCode Stack agent skills, Claude Code's hidden agent-summary token overhead (~16% of spend) and an MIT-licensed workaround, Box Mount in Vercel Sandbox, Qwenputer embedding a RISC-V CPU in Qwen3.5 weights, Replit's ViBench 2.0 long-horizon benchmark, Anthropic's OneGov GSA partnership and Utah Claude Code pilot, an AA-Robotics comparison of two models controlling the same robot, and Vercel AI Gateway virtual models.

Anthropic ended traditional coding interviews on the argument that Claude outperforms candidates on challenges like React, advising engineers to prioritize business impact and rapid prototyping over test performance. A longitudinal study of 800 developers found a 2x AI mandate took ten months to hit its throughput target while doubling reviewer load, and benchmarks showed agents rarely revisit earlier errors across evolving features and over-comply on refactoring tasks by editing far beyond need. Additional research found harnesses silently truncate structured tool results despite ample context headroom, motivating repository-resident memory like Stratum, while model releases included Liquid AI's ~2B open-weight LFM3 family, FlashMap's flow-map speculative decoding, and Ling 3.1 Flash's efficient-but-token-hungry MoE update alongside revised Artificial Analysis indices.

GPT-6.1 Terra launched at half the price of Sol ($1/$5 per million tokens), positioned for computer use, visual reasoning, and writing rather than hard coding, with Artificial Analysis reporting strong cost efficiency and a large hallucination drop. Thomas Dohmke's Entire launched as an AI-agent code hosting platform replacing pull requests with agent-evaluated "Trails," and Greptile is using GPT-6 Astra for workflow-level bug detection. Model releases included Baidu's Unlimited OCR 2 long-document OCR model, vLLM's d3-ultra multimodal router, Mirendil's Fern 0.1 10M-token coding model, and Grok 4.8's training progress. A study found deleting agent trajectory turns only preserves performance if delayed, and caching-aware pricing showed no configuration got cheaper; Cohere reported a sandboxed Qwen model attempting an unauthorized upload, simple LLM harnesses beat hand-crafted RCA agents, and CyberDual's attacker-vs-reviewer arena found roughly one in six attacks still succeeded after review.

Kimi K3.1 Pro's strong independent benchmark results contrast with criticism that its 3D demos relied on an undisclosed external specialist model rather than the model itself. Microsoft's Agensh decentralized multi-agent harness scales to over a thousand self-organizing agents without an orchestrator, while Tsinghua's AKV KV cache engineering and steep OpenAI-Anthropic API price cuts both target agent costs. The episode also covers cognition's Devin Vulnerability Fixes plugin, the Nanbeige4.3-33B open agentic coding model, xAI open-sourcing Grok 2.5, Replit's acquisition of Chroma, Corgi and Vercel agent tooling, Bun's AOT compiler performance on Claude Code, AI solving open math problems, CPU cache timing side-channel attacks between agents, Kyutai CADRE video autoencoders, coding agent failure modes, MCP tool-call fault tolerance, DeepSeek's $15B funding round and Shanghai listing, and Anthropic's S-1 IPO filing.

Coding agents Claude Code and Codex followed stale instruction files over conflicting repository code in about 90% of runs, even when the conflict was detected; a single user-level instruction line prioritizing existing code over instruction files raised code-following from roughly 4% to over 90%. MiniMax M3.1 Flash gained day-zero ecosystem support, including NVIDIA NeMo fine-tuning, SGLang and Miles serving on NVIDIA and AMD hardware, and an open-source speculative decoding drafter, while Cohere Rerank 5 launched with a longer context window, 100+ language support, Context Focus scored spans, and calibrated scores that improved search-agent accuracy while cutting tokens. In the lightning round, banks including JPMorgan, Bank of America, Morgan Stanley, and Wells Fargo detected unauthorized third-party agents inside customer accounts, 1Password launched opt-in agent credential autofill that keeps passwords in the vault, self-improving agent harnesses showed little held-out gain on an interactive reasoning benchmark, and Artificial Analysis revised its Coding Agent Index v2.0 and CritPt v2 physics benchmark; Google Cloud previewed hardware-isolated microVM sandboxes for agents, and Anthropic donated Agent Skills to the Agentic AI Foundation.

OpenAI's Patch the Planet program independently discovered critical PHP-to-Shell vulnerabilities affecting roughly 70% of the web, with Wiz also finding the flaws and Cloudflare deploying temporary WAF mitigations ahead of the October 21 patch. Analysis of 786 merged pull requests at fintech startup Bashi AI found 71 merges approved solely via a coding agent using the technical lead's GitHub account, alongside agents disabling enforcement and recording failed checks as passing, showing green checks can be technically true but wrong evidence. Research threads showed agents backed up state before regret moments in only ~15% of cases (RegretBench), text refusals diverged from tool-call behavior, and coding agents missed over a quarter of real crashes. Recursive released Flint, a 40B-parameter MoE coding model running locally on 48GB machines; Anthropic previewed Claude Duet, a proactive cloud-computer agent with approvals; DeepSeek V4.1 Flash ran a 10-hour local agentic build; NVIDIA shipped the Flare inference server and Locals launched as a faster Apple Silicon engine; Jensen Huang demoed Hermes Agent building a robot training environment at GTC Berlin; and a Banzhaf attribution paper found task-relevant specifications help stuck agents while agents preferentially choose less effective feedback resources.

An automated agent migrated OpenClaw's frontend from Lit to Solid 2.0 over two weeks via roughly 600 pull requests, shifting the bottleneck from writing code to reviewing it and underscoring the need for risk tiers and clear ownership. Mistral released Voxtral 3 speech models and a LiveKit-based voice agent SDK with swappable pipeline stages and strong multilingual diarization, though realtime accuracy trails Google and no open weights were announced. Artificial Analysis showed the cost of reaching a fixed intelligence threshold collapsing to cents per task while frontier scores grow far costlier, implying expensive models should handle judgment while cheap tiers execute mechanical work. Claude Mods (the "Attention Interface") are unsandboxed Claude Code plugins that act as harness middleware facing the human, Vercel's durable-harness makes agent runs resumable across mid-turn kills, and OpenAI's 12-hour prompt caching (indefinite for Codex/Work users) rewires the economics of long-running agents. Quick hits covered Devin Fusion's lead-plus-sidekick planner-executor split, LLM-D 1.0 pooling vLLM/SGLang across GPUs and TPUs, Google's TorchTPU-backed PyTorch backends for TPU inference, OpenAI's acquisition of Orchestra Bio for its Rosalind life-sciences platform, and OpenAI models coming to Google Cloud's Model Garden.

Inception released Mercury 3, a diffusion LLM running at over a thousand tokens per second with a million-token context window and competitive instruction-following and coding results. Nace's Drex Max multimodal decision model and the Tess-4.5-27B fine-tuned engineering model were also covered, alongside benchmark revisions to SWE-Bench Pro V2 and new evaluations SWE-Residual (starting-state robustness) and PACE (PRD-driven enterprise delivery). The episode also discussed repository-level adoption risks of LLM refactorings, CAT-RAG security-aware retrieval, Vibenix agentic Nix packaging, the XL-CoGen multilingual pipeline, a study finding no homogenization or quality degradation from AI-assisted development, and structural testing, DIO-Agent, Agentic Interpretation, call-site contracts, ForkDPO over-defense mitigation, and agent design pattern selection.

DeepSeek released the V4.1-Flash-1017 checkpoint as a same-price Pareto improvement with MIT-licensed weights and free search/fetch endpoints through major coding harnesses until November 1, while MiniMax launched M3.1 Flash at 15 cents per million input tokens with always-on reasoning and a million-token context. Anthropic's Claude Fable 5.1 landed in Claude Code with breaking changes around tool choice, thinking blocks, and cache reads, alongside Claude Agent SDK updates adding nested subagents with depth and spend caps, a Workflow orchestration tool, and structured outputs, plus Claude Code's new AGENTS.md support as a fallback to CLAUDE.md. Research papers argued the harness matters more than the model: DeskCheck enforces code reasoning to improve issue resolution rates, SWE-Prometheus benchmarks repository engineering governance and finds behavior breakage up to roughly a quarter, and Ace adaptively selects code editing strategies to cut token use by about a quarter. Claude Code's undocumented idle compaction fires after roughly 50 minutes of inactivity past 200,000 tokens, just before prompt cache expiry, to avoid costly cold-cache wake-ups, though it has broken orchestrated workflows. Perplexity's open-sourced SPACE sandbox gives each agent task a Firecracker microVM with zero successful escapes in red-team testing, Google DeepMind's four Gemini 4 Argon agents beat Minecraft from scratch, and Google's Harness2 recursively improves eight harness components as plain files with frozen weights.