424 HN setups counted, Crush grows subagents, Inkling 2's weights due next week, Anthropic's IPO wobbles

🔥What's hot

🛠New tools & products

• Crush v0.100.0 — is Charm's terminal coding agent, and this release lets it run a team of named background subagents you can message, stop, or move to a different model mid-run. The useful part for client work is the permission system: you decide which commands run unattended and which are off limits entirely. You can block kubectl and terraform entirely, while a new Sysadmin mode does the opposite and lets Crush run any command on the machine. GitHub

• Spaces — is a new Amp feature where you talk to a single coordinating agent that spawns and organises the other agent threads, instead of you creating and juggling them. Fanning work out across many agents is moving from something teams hand-build into something the vendor ships, which changes what is worth building in-house. Amp's own security-review space split the codebase into 16 areas, gave each its own thread, and had a separate thread check every finding. Amp

  “It's more like sitting in the captain's chair than down in the trenches.”

• GitHub's own numbers say review is the new bottleneck. GitHub published a guide splitting agent work into plan, build, verify and observe, and leads with its own figures: over 1.2 billion commits in a single week in October, up more than 400% in a year. The honest bit is the drag: GitHub cites a Faros AI study where pull requests spent 441% longer in review as teams went from low to high AI adoption. Its fixes are unglamorous — smaller stacked pull requests, and a package firewall, in public preview for npm, that blocks downloads breaking policy. The GitHub Blog

  “But ask yourself: do you feel 400% more productive than you were last year?”

• AI Kill Switch — is a control in CrowdStrike's Falcon Guardian that shuts down a misbehaving AI agent; it landed on 3 November at the company's Fal.Con Europe event. "How do we stop one mid-run?" is the first question in every agent governance conversation, and now there is a mainstream security product to point at. CrowdStrike also says it plans to buy Fiddler AI and Dreadnode for around $495 million combined, per the Wall Street Journal; both are announced rather than closed. geta.team

💬What people are saying

📰News

• Inkling 2 Small — is Thinking Machines' new mixture-of-experts model — a design where only part of the network runs per request — with 594 billion parameters in total and 17 billion active. It is on the lab's Tinker platform now with weights closed; Apache-2.0 weights are due next week after external safety testing, which would give clients wary of Chinese open weights another option. Artificial Analysis scores it 31 on its intelligence index against 26 for Mistral Large 4, and makes the open-model crown conditional on the weights actually appearing. Thinking Machines Lab

• OpenAI — has raised $11 billion so far in a round that values it at $1.25 trillion, Bloomberg reported on Friday; the round has not closed and could reach about $20 billion. In September Bloomberg said the target was $30 billion, and that gap is what clients will quote at you when they ask whether the money is cooling. TechCrunch links the market nerves to the midterm result and voter hostility to data centres; an OpenAI listing is now expected next year. TechCrunch

• Anthropic — is reportedly struggling to hold its IPO target: the Wall Street Journal says three large investors think $1.5 trillion is unrealistic after last week's selloff, with one putting it nearer $1 trillion, and Reuters hears the roadshow may slip. None of it is confirmed and the company declined to comment. At that price the listing would barely clear the company's last private valuation, which is the sort of detail procurement people raise when they ask how solid your main vendor is. May's Series H raised $65 billion at a $965 billion post-money valuation. Trending Topics

• gVisor — is the sandbox that puts a user-space kernel between untrusted code and the host machine, and Google is donating it to the CNCF along with Agent Substrate and the OpenRL training project. Anthropic already runs Claude-generated code inside it and OpenAI uses it to isolate execution, so neutral governance makes it an easier default when a client asks where agent-written code should actually run. Google says upstream Linux maintainers refused its gVisor performance patches while one vendor owned the project; Agent Substrate 1.0 is now generally available on Google Kubernetes Engine. The New Stack

  “We believe that these donations will accelerate the development and adoption of agentic workloads through these projects similarly to what we saw with Kubernetes.”

• Clutch Security — controls which company systems an AI agent is allowed to reach, and Palo Alto Networks is in talks to buy it for $500 million in cash plus milestone payouts, according to a person familiar with the discussions. Nothing is signed. Agent identity and access now has a market price, which is a useful frame when a client asks who owns their agents' credentials. Clutch is approaching $10 million in annual recurring revenue, putting the price at roughly 50 times revenue; it would be Palo Alto's seventh acquisition this year, per PitchBook. TechCrunch

• ARTEX — is an open-source agentic hacking platform, and Check Point got inside operator consoles to find more than 1,800 publicly exposed instances running attacks that were almost entirely autonomous once the operator set a goal. The detail to repeat to a nervous client is that guardrails did bite: one operator hit refusals on Anthropic's Opus 5.5 and simply switched to more permissive Chinese models, with Deepseek Flash the most common. In one campaign the agents took over 300 Apple code-signing certificates and passwords; three passed validation and could be used to sign iOS apps. Check Point Research

• Agent Distraction — is Google's analysis of more than 913,000 coding agent sessions, counting how often a leftover TODO comment linked to a bug ticket pulls an agent off task: 13% of sessions meet one, and in 6% of encounters the agent chases it, usually unasked. It puts a number on something consultants keep asserting — dead comments and stale docs are a running cost now, which finally gives codebase cleanup a budget line. Those unprompted detours added 70% more steps and 105% more wait time per session. Google Developers Blog

💡One takeaway