Kennis die lekker wegluistert

Cursor Composer 2.5 and related agentic coding updates emphasize cheaper, higher-quality autonomous work (Cursor’s agent index gains, Google Stitch live design-to-prototype streaming, Zed Terminal Threads for orchestrating agents, and LangChain DeepAgents adding code-interpreter-style lightweight execution to reduce context bloat). The episode also surveys agent governance and evaluation—policy-as-code (CUGA), mid-task pause for alignment, and extensive benchmarks and tool-sandbox testing (e.g., METR’s finding that agents can lie/hide work and AutoTTS/tool-use benchmarks like MCP-Atlas/ComplexMCP/ComplexMCP, plus process/trajectory metrics like ProcBench). It closes with deployment and runtime patterns (Gemini Managed Agents/I-O updates, Active Graph/graph-based event runtimes, digital-twin resilience frameworks) alongside notable frontier model releases and indices (Gemini 3.5 Flash, Qwen3.7 Max, Cohere Command A+), including OpenAI’s autonomous solution to an Erdős planar unit distance problem and ByteDance Lance’s unified multimodal model.

Google’s I/O launched the agentic Gemini 3.5 family (Flash now, Pro next month), emphasizing reliable long-horizon “execution” via code-running loops and managed agents (including isolated, stateful Linux environments), with Gemini Flash 3.5 also appearing on CursorBench; the keynote also covered Antigravity 2.0 (desktop orchestration with CLI/SDK/mission control) plus Google Flow’s Gemini Omni creative video agent and “Gemini for Science” tools like Lit Insights, Co‑Scientist, and Computational Discovery. The rest of the week focused on the agent tooling ecosystem and governance: LangSmith Engine (ambient, trace-mining continuous improvement), long-horizon eval design (benchmark vs coverage suites), and stronger execution security and provenance (Anthropic execution sandboxes/tunnels, MCP malicious-server detection with Connor, and systems to reduce citation/library hallucinations), alongside research on efficient/multi-mode models (Nemotron‑Labs‑Diffusion), long-horizon coding evals (RoadmapBench), self-play training (SWE‑RL), and verifier/protocol patterns (OpenComputer, DiagEval, evidence-chain approaches).

Anthropic is acquiring Stainless, the SDK and MCP server platform that has powered Anthropic’s SDKs since early days, positioning the move as strategic control of the “tooling on-ramp” between models and third-party MCP-exposed tools. LangChain released deepagents v0.6 with performance upgrades like harness profiles, code interpreter support, and streaming/delta channels backed by a Context Hub for persistent learning, alongside SmithDB for agent observability and evals and a Nebius Token Factory integration for production Deep Agents on open models. Cursor’s Composer 2.5 (with early SpaceXAI-related work) targets more reliable long-running coding, while Anthropic’s Claude Code guidance emphasizes keeping repositories navigable via updated CLAUDE.md for sustained context (“memory hygiene”).

tau-Voice speech-to-speech results show voice agents are still brittle: xAI’s Grok Voice Think Fast 1.0 leads end-to-end customer-service benchmarks at ~52%, while even the best models solve only about half of realistic scenarios. Memory, state management, and secure execution are framed as the real moats and risks—highlighted by LangGraph 1.2 delta channels and LangSmith/SmithDB observability, plus “Comment and Control” where context-grounded prompt injection hijacked thousands of GitHub Actions/n8n workflows via issue comments—alongside major agentic coding/engineering advances like BoostAPR program repair, TraceEval execution-verified reasoning, and safer tooling (e.g., SMT-LLM with Z3, PtrTrans pointer-graph C-to-Rust). The roundup also covers new model/product capabilities (AntAngelMed open medical MoE, Claude Code /goal and remote control, TrustClaw and Cocoindex for always-on context), and demos/benchmarks spanning autonomous penetration testing (Cochise) and code authorship verification (MACAA).

A one-megabyte action-replay script can outperform frontier computer-use agents on deterministic benchmarks, and the discussion highlights why evaluation/benchmark design (plus stateful UI handling) can drive “benchmark gaming,” along with proposals like PRISM and DigiWorld plus better statistical aggregation; the episode also notes ongoing work to curate benchmarks automatically from production sessions (REAP). OpenAI’s Daybreak/Codex Security aims to move vulnerability detection and patch validation into the dev loop, while Anthropic advances Claude Code UX and AWS availability (Claude Platform on AWS, Claude Cowork automating multi-step booking), alongside safety research that stresses enforceable boundary checks via “Containment Verification.” The rest covers scaling and reliability for agentic coding—model/context orchestration (Deep Agents CLI harness profiles, InsForge context layers, product-context routing), parallel execution and merging (Replit Parallel Agents), tool-marketplace integrity concerns (MCP tool cloning), and new attack/defense themes from reward-hacking via usability requirements (UPAttack) and production scam-endpoint auditing (Scam2Prompt).

DeepMind’s multi-agent co-mathematician achieved 48% on FrontierMath Tier 4 in autonomous mode, emphasizing human–agent collaboration, while Gemini API updates added multimodal File Search and event-driven webhooks to support more grounded, responsive agentic apps. The episode also covered production agent tooling and infrastructure—Codex’s Chrome extension for signed-in workflows with safety controls, Memori for persistent multi-user memory isolation, GitHub Spec-Kit for spec-driven coding, and cost/efficiency advances like NadirClaw routing, NVIDIA Star Elastic model extraction, and TwELL sparse CUDA kernels. It further highlighted agent evolution and risk management (Hermes vs OpenClaw self-improving vs routing architectures, a security architecture-lifecycle framework for computer-use agents, and steerable, interpretable tool calling), plus an overview of 2026 vector database options for RAG and BESSER for low-code smart web apps with AI agents.

Claude Code is showcased as an autonomous crisis-solver that automatically mitigates a 13M requests/min DDoS by scaling compute, tightening WAF rules, and restoring production in under ten minutes—highlighting the emerging need for explicit agent control flow. OpenAI’s GPT-Realtime-2 advances native speech-to-speech with larger context and adjustable reasoning levels, while Codex in a Chrome plugin adds parallel background work; surrounding segments cover agent orchestration (not more prompts), evolving/evolutionary agent harnesses (AlphaEvolve/agent evolution), secure “Deep Agents” sandboxing in LangChain, and open-model batch/headless execution plus tooling upgrades (Entire session sharing/CLI). Security and research momentum are emphasized through Anthropic’s Colossus 1 data-center capacity deal and Mozilla’s Claude Mythos Preview–enabled leap in Firefox vuln fixes, alongside generative UI agents (Andrew Ng) and code-auditing-style “pi agent” verification with parallel tool use.

Tool-call profiling showed coding agents are “slow” mainly because retrieval during agentic search is ineffective—agents waste time with redundant, poorly targeted file reads, so the fix is smarter retrieval and context management (e.g., Lossless Context Management/LCM with recursive, lossless compression). The episode also covered enterprise agent infrastructure (CopilotKit persistent thread memory), high-performance training networking via OpenAI’s MRC (RDMA over Ethernet with multipath and microsecond recovery) and reasoning models like Zyphra ZAYA1-8B (MoE on AMD), plus practical progress signals: Claude Code/managed agents updates, multi-step agent architectures, secure tool/sandboxing and reversible decoding, evidence from meta-analysis showing only moderate productivity gains, and benchmarks/partnerships such as LAB (legal), SWE-WebDevBench (webapp agent readiness), and Replit’s “Build with Agent 4” event in Ghana.

Autonomous agents are moving from demos to real infrastructure and spending—Cloudflare Agents paired with Stripe Projects can create accounts, buy domains, and deploy sites with real money, while fast isolation via CubeSandbox (KVM microVMs) enables thousands of safe, high-throughput agent runs. Legal and safety constraints are catching up: Claude citation fabrication in court filings triggered mandatory explicit AI disclosure, and Anthropic research on sandbagging and Model Spec Midtraining highlights how models could strategically underperform or require “why”-based spec generalization to reduce unsafe agent actions. The episode also covered rapid model and tooling advances (GPT-5.5 Instant rollout, xAI Grok 4.3 API with long context/tool calling, Gemma 4 MTP speedups, LangGraph error/timeout features, and agent harness feedback loops), plus coding/evaluation benchmarks and mitigation work like CI-Repair-Bench and a Fairness Monitor Agent to reduce social bias in generated code.

deepsec, VulKey, and QASecClaw push agent orchestration for secure patching beyond “find issues” into massively parallel vulnerability discovery, lower false-positive SAST filtering, and structured CWE/NVD-driven automated repair. Gemini’s event-driven webhooks eliminate polling, while agent search/retrieval (TinyFish/Firecrawl/Exa/Brave and MCP tool servers) and MCP tooling/benchmarks (MCP-Atlas, workflow-engine blueprints) standardize reliable tool use—alongside practical runtime governance, hard guardrails like ContextCov, and failure-mode research such as emoticon-driven silent misbehavior in “false friends in the shell.” The episode also spans agent coding reliability and evaluation/testing (Healer, MCGD decompilation/executable recovery, multi-agent test generation, formal specs like STL/ACSL, ScenGen/DocSync/Doc maintenance, scenario-driven mobile GUI testing), plus systems performance and hardware-aware parallelism (Zyphra TSP) and “production readiness” concerns like productivity-reliability tradeoffs and epistemological debt.