AgentStack Daily

OpenClaw v2026.6.6 ships, Anthropic responds to a US directive suspending Fable 5 and Mythos 5 access, and an AI agent bankrupts its operator during a DN42 network scan. Other items: a coding agent damages Fedora and Linux systems, a macOS local-agent walkthrough trends on Hacker News, Claude Desktop spawns a 1.8 GB Hyper-V VM on every launch, Anthropic model naming strings get decoded, Apache Burr debuts as a reliability-first agent framework, Hugging Face's open-r1 reproduces DeepSeek-R1, and DeepSeek's own notes pull in 205 points of Hacker News discussion.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-69/

OpenClaw v2026.6.5 and OpenAI Codex rust-v0.139.0 ship this week, paired with Anthropic's Claude Fable 5 and Mythos 5 announcement and system card. We unpack the silent application risks in Claude Fable, Apple's Gemini-based AI architecture reveal, and DeepSeek V4 Pro's claimed precision edge over GPT-5.5 Pro. Plus: OpenAI's confidential draft S-1, GPT-2's 2019 staged release as a lens on today's debates, AWS Bedrock's data sharing requirement for Mythos, and a paper asking if grep alone is enough for agentic search.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-68/

OpenAI Codex ships rust-v0.138.0 with CLI-to-desktop handoff and local image path exposure; Claude Code CLI hits 2.1.169; the MCP July 2026 Release Candidate goes stateless with an extensions framework; Apple WWDC delivers a real Siri AI overhaul with Gemini and natural-language Shortcuts; and Alibaba's Qwen3.7-Max and Qwen3.7-Plus enter the agent model race with 1M-token contexts and multimodal support.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-67/

OpenClaw v2026.6.5-beta.2 and Claude Code 2.1.168 lead the agent-harness cycle, and the cycle opens with a Friday June 5 outage that hit Claude API, Claude Code, claude.ai, and Claude Cowork for roughly two hours — primarily Opus 4.7 and 4.8 — peaking near a thousand Downdetector reports. OpenClaw switched release trains to a monthly patch cadence with the June 2026 floor at 5.28. Claude Code shipped a focused day-late bug-fix release on the .167 baseline, closing session attachment, stream-json event ordering, and interrupt handling regressions that some users reported during the outage window. OpenAI is reportedly planning its biggest ChatGPT overhaul yet — a unified superapp that folds in Codex, agents, and third-party services ahead of a fall IPO. Apple WWDC 2026 opens June 8 with a Gemini-powered Siri as the headline. Anthropic expands Project Glasswing to 150+ organizations and signals Mythos-class capabilities are coming in weeks. Microsoft launches MAI-Thinking-1 and MAI-Code-1-Flash into GitHub Copilot. Gemma 4 12B ships an encoder-free multimodal design for 16GB local Macs. The MCP lane is brief this week — a one-paragraph blip, not a deep-dive. Project radar covers A2A v1.0 and the CheetahClaws Python harness.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-66/

Hermes Agent v0.16.0 — "The Surface Release" — ships a real native desktop app with OAuth remote connect, drag-and-drop file input, and a browser-based admin panel. Codex 0.137 adds multi-agent v2 runtime choice persistence and parallel web search. Claude Code 2.1.166/2.1.167 introduces fallback model chains and glob tool-name deny rules. Gemma 4 12B is Google's latest open-weight 12B model that runs locally on a laptop with 16GB VRAM. The project radar covers the A2A protocol hitting v1.0, Kimi Code CLI as a TypeScript-native terminal coding agent, and the awesome-ai-agents-2026 curated resource list.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-65/

Claude Code 2.1.165 is the latest npm `latest` as of June 5, following 2.1.163 and 2.1.164 — all bug-fix and reliability releases that clean up background sessions, plugin hooks, skill syntax, and Windows path handling. Microsoft dropped a seven-model MAI family at Build 2026 on June 2, with MAI-Code-1-Flash as the headline: a 5B-parameter coding model trained on GitHub Copilot production harnesses, scoring 51% on SWE-Bench Pro and 60% leaner on tokens than comparable models. The episode also covers the GitHub Project Radar around agent memory, code graphs, and MCP tooling that serve the local coding-agent stack.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-64/

[00:00] Episode hook OpenClaw v2026.6.1, Hermes Agent v2026.5.29.2, and Claude Code 2.1.162 drop in the same episode window. The stable OpenClaw tag is v2026.6.1, the Hermes stable tag stays at v2026.5.29.2, and the latest Claude Code npm `latest` is 2.1.162. OpenClaw v2026.6.1 ships Workboard orchestration, a governed Skill Workshop, SQLite-backed state recovery, and MiniMax M3 provider support. Claude Code 2.1.162 adds waiting-for visibility in `claude agents --json` and a batch of permission and interrupt fixes across five releases from 2.1.158 to 2.1.162. Qwen 3.7 Max and Plus split the coding-reasoning and multimodal-vision lanes. agentmemory makes every agent on your machine share a persistent context layer. This is a 60-minute episode — keep the existing builder stories and extend runtime.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-63/

AgentStack Daily EP062 leads with Codex `rust-v0.136.0`: better TUI diagnostics and error context, improved app-server lifecycle handling, named hooks and permission scopes, Python SDK and Node SDK improvements, and non-interactive installation support. Stanford's CS336 course publishes a formal AI agent guidelines document that reaches 1,863 stars in under 24 hours — institutional validation that agent workflow guidelines are becoming a first-class engineering concern. OpenAI puts GPT-4.5, o3, and Codex on AWS Bedrock, completing the pattern where both major labs distribute through the same cloud. Expanse from YC P26 uses cluster-specific fine-tuned models to predict GPU job resource needs and outperforms frontier LLMs by 8x on that task, backed by real HPC telemetry and SLURM/Kubernetes integration. The project radar covers agent OS for hardware, terminal context managers, MCP workflow templates, and physical agent scheduling.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-62/

OpenClaw `v2026.5.28` leads with steadier agent and Codex runtime recovery, safer channel delivery, stricter browser and automation inputs, provider and media expansion, externalized plugin surfaces, and bounded release proof. MiniMax M3 follows as the model-discovery headline: an open-weight-positioned coding and agent model with MSA sparse attention, up to 1M context, native multimodality, API availability, MiniMax Code integration, and open weights promised after the launch window. Claude Code `2.1.159` is the quiet CLI hygiene lane. Then EP061 evaluates Understand Anything, agentgateway, MCPJungle, and CodeAlmanac as practical tools for codebase graphs, agent traffic control, MCP management, and persistent repo context.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-61/

Claude Code `2.1.158` on npm `latest` adds auto mode for Bedrock, Vertex, and Foundry when `CLAUDE_CODE_ENABLE_AUTO_MODE=1` is set. OpenAI's May 29 Codex update adds Windows computer use in the Codex app, remote control from mobile or Mac while the Windows host keeps project files and local context, in-app browser infrastructure improvements, and Codex Profiles for identity, activity, usage stats, and token activity. The episode then covers Anthropic's Messages API system entries as a runtime-control surface, plus OpenLore, Mnemo, OpenMonoAgent, and Prometheus as project-radar examples of local architectural memory, persistent agent cognition, local-only coding agents, and graph-backed repair loops.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-60/