In partnership with
Good morning. Here's what's worth acting on today.
BREAKING STORY
Google, OpenAI, Microsoft and Anthropic back one robots.txt token for opting out of AI training
On 30 October, four AI companies backed a single robots.txt token, __usage_ai-training, that tells their crawlers whether your pages may be used to train models. It comes from an individual IETF draft, "Machine-Readable Training Preference Signal", written by Anthropic's Kevin Kelley. It is not an IETF standard, and it covers training only, not crawling.
Google, Microsoft, Anthropic and OpenAI back it, with Meta supportive but not committed. OpenAI says support comes "in the coming months" and Google "sometime next year". Microsoft says new preferences take effect from 1 February 2027 as pages are crawled, and it adds a separate AI Grounding control for whether a page can be used in answers at query time.
My take: Block training on your paid material: courses, templates, paid posts. Keep your free pages open so AI answers can cite them. At Microsoft, at least, training and answer-time use are now separate decisions.
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
MARKET TRENDS
AI spend per engineer is up 3.8x since Q1, with headcount flat
Companies are paying far more for AI per engineer without hiring more of them. Jellyfish's own data shows the median company spent 3.8x as much on AI per engineer in the third quarter of 2026 as in the first, while ending the quarter at 98% of its first-quarter headcount. The median developer used 59M tokens a week in September, about 8x a year earlier.
TechCrunch, citing Jellyfish data on 276,000 customers, reports a narrower figure: spending on Claude Code and other tools built on Anthropic models has grown 10x since the start of the year, with their cost now about 2% of total engineering cost.
How to make money from it: AI spend is a new line that most small teams never budgeted for. Sell dev shops and agencies a fixed-price AI spend review: audit their token use, route routine tasks to cheaper models and add caching. If you build with agents yourself, track your own agent cost per task and keep it visible.
GROWTH HACK / DIY
Bring your own plan: let users pay for the AI with a subscription or key they already have
The idea: Your users already pay for AI, so stop charging them for it twice. v0 now lets ChatGPT Plus and Pro subscribers connect their plan (docs updated 30 October): eligible requests count against their ChatGPT limits, v0 charges a small platform fee, and when the ChatGPT limit runs out it pauses and asks rather than switching to v0 credits. The ChatGPT connection is open only to open-source partners and selected private clients, but the API-key version is one any solo builder can ship this week.
Do it today:
Split your price on paper into platform (workflow, storage, support) and model (tokens). The platform fee is what you charge whoever's key pays for the tokens.
Add a "Use your own key" setting. Validate each key with the cheapest possible call, encrypt it at rest, and route requests to the user's key when they have a valid one.
Pause at their limit and never fall back silently. Save the work and offer two buttons: try again later, or continue on your credits.
Illustrative maths (made-up numbers): on a $20 plan where an average user's model calls cost you $12, you keep $8, and a heavy user costing $35 loses you $15. A BYO-key plan at $8 a month keeps the same $8 margin on every user, light or heavy.
What you'll need: your app, a secrets store or encrypted database column, and your billing provider. Optionally OpenRouter's BYOK, which charges 5% of its normal price after a monthly free allowance (as of 31 October 2026). Time: a weekend for one provider.
AGENT LAB
Async subagents: background workers that push results back instead of being polled
The orchestrator starts a subagent on its own thread, gets a task ID back straight away and keeps talking to the user. When the worker finishes, it notifies the orchestrator itself. LangChain shipped this in Managed Deep Agents v0.10 on 29 October. On a 52-slide reskin, its design agent ran for about an hour while its subagents worked more than six hours between them, it says. Use it for long jobs and fan-out; skip it for quick lookups and for work that needs its own sandbox.
How it works: start task → task ID returned → worker writes only the files it owns and drops its result in an inbox → delivery depends on the orchestrator's state. Busy: a check before its next model call picks the result up. Idle: the result wakes it with a new run. Waiting on a user approval: the result queues until the user answers.
Try it: Build a research-brief bot with three async workers (competitors, pricing pages, customer complaints), each owning one file. Then break it on purpose: kill a worker, push a duplicate result, and point two workers at one file. Each should fail loudly.
TRENDING AI RESEARCH
The best vibe coding agent meets 70% of what you ask, but fully finishes only 26.4% of the edits
Co-Design Bench took 273 editing requests from real coding sessions and graded ten model endpoints against 2,331 hidden acceptance criteria. The best of the ten models met 69.9% of requirements per task on average but fully completed only 26.4% of tasks. On the repository track, every model scored lower on requests that bundled three or four types of change than on those with one or two.
Why it matters: Averages flatter agents. A 26.4% full-completion rate means roughly three in four edit requests needed another round from you. When your agent says "done", treat it as about 70% done. The authors say the bundling result is descriptive and does not prove cause.
How you can use it: Write your acceptance checklist before you prompt, and tick it in the browser, not in the code. Ask for one type of change at a time: structure, style, interaction or content. Then open a fresh session and ask it to prove each ticked item is not done, hunting for stubs, hardcoded values and features you can't reach.
TOOL REVIEW
Faylee: bug reports your coding agent can read and fix from one link
What it does: A Chromium extension records your screen, console, network requests and clicks on one timeline and turns the recording into a link. Its MCP server lets Claude Code, Cursor, Codex and others open that link, inspect the failing requests and comment back; Faylee says it strips secrets in the browser before upload. It launched on Product Hunt on 30 October.
Price (as of 31 October 2026): Free for up to 3 members, 30 recordings a month, 5-minute videos, 14-day retention and the MCP server. Team costs $12 per member a month billed yearly, or $15 monthly. Recording links for people without Faylee and the on-site bug widget are paid-only.
Who it's for: Solo builders and small teams shipping web apps with Claude Code, Cursor or Codex. It does not yet suit native mobile apps.
Verdict: Worth installing if you build with Claude Code or Cursor; the free tier covers your own bug-hunting, pay only when customers need to report bugs.
What Got Funded Yesterday
Five newly funded startups: three announced on 30 October and two on 29 October.
PhotonPay (photonpay.com) · fintech · Hong Kong. Founded by Lewison Chen, PhotonPay gives businesses one API for wallets, cross-border payments, FX and card issuing. It says stablecoins sit at the core of its system, backed by 34 licences and US$65B in annualised volume. It raised a US$40M Series C led by Blockchain Capital, with SBI Group, IDG Capital, GL Ventures and Lightspeed Faction.
Ares Intelligence (aresintelligence.ai) · AI software · UK. Founders Matt Falconer and Alfie Ranstead built an AI interviewer that holds private voice chats with employees, in several languages, to find where AI could help. It says it turns hundreds of these talks into a ranked, evidence-based AI roadmap within weeks. It raised €2.6M from Creandum and Lakestar, with angels linked to Google DeepMind; no lead was named.
Oxus Metals AI · AI mineral exploration · Singapore. Founder David Zhen Yin, a Stanford Mineral-X co-founder, built an AI that combines exploration data to rank drill targets for critical minerals. It says it tests many subsurface models rather than betting on one geological theory, and partners with asset owners in joint ventures. It raised nearly US$36M in an initial round at a reported US$180M valuation; it named no investors.
KineDB (kinedb.com) · developer tools · Nordics · announced 29 Oct. Co-founders Axel Löwenhoff and Allan Zimmermann are building one database that runs both in the cloud and on devices with no internet, then syncs when they reconnect. It says a single engine replaces separate offline and backend systems. It raised €850K in pre-seed led by Superhero Capital, with NGL Ventures and angels.
Fronted (fronted.com) · HR tech · Norway · announced 29 Oct. Co-founded by Marius Blaker Høgevold (CEO) and Hans Nyvold Kjellby (CCO), Fronted lets companies recruit abroad, employ through its Employer of Record service and set up local entities from one platform. It says it builds local labour rules into the software to keep pricing clear. It raised a €1.7M seed led by Kaknäs Invest, with Skagerak Capital, Antler and Sondo Capital.
QUICK HITS
If you installed or updated the Claude Code extension in Cursor on 29 October, check your machine: tampered builds (2.1.312 and 2.1.313) in Cursor's marketplace tried to add an SSH key and open a reverse tunnel, and Cursor has published remediation scripts.
ElevenLabs released SFX v3, a sound-effects model that follows prompts asking for several sounds at once; it is live in ElevenCreative and the API.
Claude Enterprise now has smart reports, in-app surveys and pooled budgets, showing what teams use Claude for and what it costs.
That's today. Hit reply and tell me what you're building. I read every email.
Dev
