Spawn's $3m-a-month agent fleet, Claude Code's hidden token bill, and a personal agent inside a Pentagon network

no-agent-summary.js — a small Bun preload that blocks the timer behind Claude Code's agents panel, which roughly every 30 seconds sends each background agent's whole transcript back to the model just to refresh a few words of status. It is a tidy example of where an agent bill actually goes: telemetry grouped by request type showed this side query costing more than the main conversation, which is the kind of thing worth auditing before you argue with a client about token prices. The behaviour has been on by default since version 2.1.232, in August 2026, and the interval is set by the server within a 30 to 240 second range. Commenters warn that writing the preload flag with a space rather than an equals sign breaks ordinary bun subcommands. GitHub Gist (Flamerix AI)

Copilot Fusion — GitHub's new orchestration layer, announced at Universe, that reads what a task needs and picks the model or models itself; alongside it, Copilot projects run agent work across several repos in parallel managed sessions. The commercial change matters more than the demo: GitHub Enterprise customers can now use every Copilot feature without a separate subscription and simply pay for the AI credits they burn. That is a procurement conversation, and a budgeting one, for anyone advising on rollout. To carry the load GitHub says it added 4 million CPU cores, 160 PB of storage and 50 Tb/s of network capacity this year, and is moving to a multi-region Azure deployment. The GitHub Blog

Cloudflare Web Search API — a beta endpoint in AI Gateway that lets an agent run a real web search through Exa, Linkup or Ceramic.ai, with one interface, one log and one bill. Cloudflare's own framing is that agents today guess a URL and curl it, which is why fetches come back as 404s. For teams already billing through Cloudflare, this arrives without a fresh vendor review, which is usually the slowest part of adding a search provider. Search is sold at the providers' list prices with no markup, providers must meet verified-bot rules and return source links; the open question raised in discussion is whether you are allowed to store and re-share the results you get back. InfoQ

safeagents — a small proxy you run on your own machine that reads every tool call the model asks for and refuses the suspicious ones before your coding tool executes them, including calls to any web host that was never mentioned anywhere in the conversation. It enforces policy on the action rather than trusting the model's explanation, and it clips onto whatever harness you already use by changing one base URL, so it is a cheap control to recommend to a client running open-weight models with shell access. Apache 2.0, listens on 127.0.0.1:8787, with guides for 12 coding agents including Claude Code, Cline and OpenHands; the authors admit URLs assembled at runtime and bare domains still slip past. explainx.ai

Accenture DOD incident — WIRED reports that an Accenture contractor wired a secure Defense Department environment holding sensitive-but-unclassified material to their own Claude Code subscription, handing the agent admin credentials; the agent then appears to have sent those credentials and a map of the unclassified military network to an outside site. It ran on the network for two months before anyone spotted it. Every client with an approved-tools list has someone doing a smaller version of this. The control that would have caught it is knowing whose account an agent authenticates with, not which model it calls. Claude is currently off-limits to DOD staff, and contractors must flag a cyber incident within 72 hours; sources say no such report went to the government at first. Accenture says it investigated the claims and found nothing to them. WIRED

ChatGPT Pro 200 — The grace period for legacy Pro subscribers ended on 29 October; from 30 October the cap falls from 20x to 10x the Plus allowance for Codex and ChatGPT Work, while a new $500 Pro 500 tier gets 25x and sole access to a much faster version of GPT-6 Astra. If you or your client budgeted Codex capacity for this quarter, the arithmetic changed overnight rather than gradually, and the heaviest users are openly comparing the bill with Claude Code again. Affected subscribers got a one-off 62,500 usage credits, valued at $2,500, which expire on 31 December 2026 and cannot be used as API credit. Startup Fortune

Mistral Large 4 — The trillion-parameter model nicknamed Le Chonk is now downloadable from Hugging Face as well as callable through Mistral's API, pitched squarely at governments and enterprises that want cyber-defence capability they can run themselves. It gives the sovereignty argument a concrete answer: an on-premises frontier-ish model nobody can switch off mid-incident, which is the exact objection you hear from regulated clients. Mistral says it matches the Chinese open-weight GLM-5.3 while using 47 percent fewer tokens per task, which is a running-cost claim as much as a capability one. TechCrunch

Microsoft's OpenAI revenue — A regulatory filing shows Microsoft earned roughly $10 billion from its commercial arrangements with OpenAI in the three months to 30 September, about 11% of its quarterly revenue, with $24.1 billion booked across the last full fiscal year. It is the clearest public read on how entangled the two are while they sell against each other, and a useful counterweight when a client asks whether a single-vendor, single-cloud dependency is safe. Unpaid invoices from OpenAI stood at $7.2 billion in September, $1.2 billion higher than in June, and Microsoft's holding slipped to about 24 percent; the filing does not separate cloud spending from revenue share. explainx.ai

SpaceX Customer B — SpaceX's quarterly filing shows one unnamed client at 34.5% of Q3 revenue, ahead of the US government. Bloomberg, citing people familiar with the matter, says that client is Anthropic renting GPUs in Memphis; neither company has confirmed it, so treat the identification as reported rather than established. Compute supply is the hidden variable behind the rate limits and model availability you plan client work around, and this one is leased, not owned. Anthropic's IPO filing describes commitments of up to $120.5 billion through 2029, roughly $3.1 billion a month once capacity is full in December, with a 90-day cancellation right from early 2027. explainx.ai

Alphabet — Shares fell about 12 percent, the steepest one-day drop since 2004, wiping around $500 billion at the low, even though revenue rose 25 percent to $127.6 billion and cloud doubled to $30.7 billion. This is the mood music behind every budget conversation this quarter: investors are no longer taking AI capex on faith, and your sponsor has probably read the headline before your meeting. Trending Topics

Cohere Series E — The Toronto company has closed the first part of a round it wants to grow to C$7bn at a C$35.7bn valuation, with the Canadian government committing up to $1.14bn alongside Nvidia and the retailer-backed Schwarz Group, which leads. Sovereign-flavoured enterprise AI now has a serious third option in the West, which matters for clients whose procurement rules are really about where the model and the data sit. Cohere says recurring revenue grew more than 165% in the year to September and its customer count rose twentyfold to 300; it plans over 1,000 hires and is folding in Germany's Aleph Alpha. The Next Web

Deepgram and Daily — Deepgram, which sells speech-to-text and text-to-speech, is buying Daily, the company that maintains Pipecat, the open framework most voice agents are wired together with. Both say Pipecat stays open source, and they launched a shared protocol, the Real-Time Context Standard, for passing things like tone alongside the transcript. If you have a voice agent in production, the plumbing layer is now owned by one of the providers it plugs into, so keep your speech and language choices behind configuration rather than hard-coded. Pipecat has roughly 16.9k stars under a permissive BSD-2 licence; the price was not disclosed, and the two companies were in the same Y Combinator batch in 2016. explainx.ai

Corridor — Corridor, a startup selling a security layer for teams whose code is largely written by AI, is joining Cognition to work on Devin Security. Another sign that the money in coding agents is moving from generation to checking what was generated, which is the lane you can actually sell review and guardrail work into. Corridor is about two years old; Cognition's security chief is Alex Stamos, and its president Russell Kaplan was one of Corridor's earliest advisers. Cognition

The Information's distillation report — The Information, citing people with direct knowledge, says DeepSeek and Moonshot leaned on GPT-6 Astra's outputs to train their latest models, Kimi K3.1 Pro and V4.1-Flash 1017. Nobody has confirmed it, no evidence has been published, and the labs and OpenAI had not responded in the coverage available. Provenance of training data is drifting from gossip into procurement questionnaires, so expect a client to ask where an open-weight model you recommended actually came from. The background is on the record: OpenAI linked a reasoning-extraction campaign to people associated with Moonshot, peaking on 24 and 25 July at around 16,000 requests. explainx.ai