

✓Founder @deptofproduct. I write about AI, trends and technology for product teams on Substack for 50k readers and 10k on YouTube.


This week, I looked at 320 open product manager job ads across 46 tech companies, and the biggest difference between AI PM roles and the rest is evals. 39% of the 31 ads with AI, agents or ML in the title mention evals, against 4% of the other 289. AI PM ads are also more than twice as likely to mention prototyping (42% against 17%) and twice as likely to mention specs. But they’re less likely to mention customer research (6% against 15%). It’s a small sample, but that gap is a little bit worrying for roles that are meant to define what good looks like for users. The full analysis is in the latest Deep Dive: departmentofproduct.substack…

Active users, latest quarter: Duolingo 64.9M daily, +29% Pinterest 661M monthly, +10% Snap 495M daily, +4% (and missed estimates) Reddit 130.3M daily, flat, with US users down 2.5M on the quarter Roblox paying users 1.8M a day, down from 2.3M Reddit expects users to fall again as it moves to logged-in only.

Three tech leaders talked about headless products this week: @HubSpot’s co-founder and CTO @dharmesh said “MCP (Model Context Protocol) is the new HTTP. If you aren’t making your software product composable and callable by AI harnesses, you’re going to miss this wave.” Days later, HubSpot announced a refounding with co-founder @bhalligan back as CEO, and Dharmesh says it’s “leaning strongly into a headless version of HubSpot”. @meetgranola’s co-founder says it now has a dedicated team for headless Granola. On the same day it launched MCP 2.0, with search that uses 85% fewer tokens and tools that let agents add agendas and organise notes. @hexanalytics CEO @barrald pointed out that the AI companies asking everyone to open up don’t always do it themselves. External MCP clients like Cursor or ChatGPT can’t connect to Claude Docs or Slides, which he summed up as “MCP for thee, but not for me”. For product teams, the pressure to make products work with agents is only going up, even if the biggest AI companies are keeping their own apps closed off. I wrote about this and more in this week’s Product Briefing: departmentofproduct.substack…

SaaS, latest quarter’s revenue growth: Pulling away Datadog +35% Klaviyo +25% Atlassian +24% (cloud +31%) Amplitude +24% ARR Toast +24% ARR Slowing down Twilio +20% HubSpot +17%, guiding to 12% for Q4 Workday +13.9% subscriptions Salesforce +6% organic

Anthropic launched its own spreadsheets, whiteboards, data stories and a meeting assistant this week, while still supplying the models behind apps like Granola, Linear and Cursor. In this week’s video briefing I looked at what that could mean longer term. If Anthropic keeps launching products that compete with the companies building on its models, some of them might move to open-source models, or refuse to connect to Claude at all to keep their relationship with their users, like Apple tends to. Most companies aren’t Apple though, and with AI making it easier than ever to switch products, refusing to integrate could easily backfire. So will the AI labs end up owning the apps too, or will the Granolas and Linears of the world hold their ground? The full video is here: youtu.be/vcaG1-hI34g

METR, the AI safety research group, assessed how much AI is speeding up AI research inside Anthropic, and found its models are likely improving about twice as fast as they did between March 2024 and October 2025. - Anthropic’s benchmark scores have accelerated 1.9x since October 2025 and METR estimates about 80% of that is down to AI, though with a lot of uncertainty. - Anthropic staff also self-reported being about 12x as productive with its most capable internal model in October. But it does also call the 12x figure “only very loose evidence”, since the answers probably came from an unrepresentative group of staff, and it says Anthropic redacted most of the evidence METR found most informative (!). Here’s METR’s full assessment which is an interesting read: metr.org/blog/2026-11-05-ant…

Anthropic has published a new guide to prompting Claude Fable 5.2. Here are some of the key principles to consider when using it: 1. Effort now does a lot of the work - at high and above, Fable 5.2 thinks before almost every reply. The default is high but for every day coding and knowledge work, medium is fine. 2. It sometimes ends a turn before the work is done, especially at low and medium effort. Anthropic includes a block of text for your system prompt that roughly halved unfinished sessions in its testing. 3. It pays closer attention to who owns the code its working on. Adding “You are the owner and long-term maintainer of this codebase” meant it consolidated duplicated code in about 75% of runs, against under 10% without it. 4. It treats any instruction with “never” in it as absolute, so old cautions in memory files and notes can pile up and make it ask for approvals it doesn’t need. Anthropic suggests using the phrase "ask for approval before deleting files" instead of "never delete files" Here’s the full guide: platform.claude.com/docs/en/…

Sources: PostHog: posthog.com/blog/self-drivin… LaunchDarkly: launchdarkly.com/blog/introd… Perplexity: x.com/AravSrinivas/status/21… Vercel: x.com/rauchg/status/21201884… AWS: aboutamazon.com/news/aws/how…

How much of the work agents now do, by companies’ own numbers: PostHog: 11% of October’s pull requests opened by its self-driving agents LaunchDarkly: 20%+ of pull requests shipped to production with no human review Klaviyo: most warehouse queries now come from its analytics agent (under 10% a year ago) Perplexity: the median employee gets 7x more agent-hours a week than in September Vercel: 14,665 vulnerabilities patched by its agent in under a week AWS: 100% of AWS Emulator’s code written by agents

How many more companies can enter the meeting notes market? 😆 ElevenAI is a new product that records meetings on your Mac, labels each speaker, writes the summary and drafts the follow-up. Dictation works in any app in 90+ languages, and it shows a notice on video calls while it’s taking notes. elevenlabs.io/elevenai

Sources: Amazon: x.com/SwamiSivasubram/status… Hex: hex.tech/blog/crashing-the-c… Datadog: x.com/datadoghq/status/21187… Zapier:

What an agent task costs now: Amazon ticket labelling: $0.0222 → about $0.0005 per ticket, after swapping Sonnet 4.6 for a decision model Hex data agent: $6.44 → $1.64 per conversation since January Datadog Bits Investigation: 84% cheaper per investigation since launch Zapier’s AutomationBench: Fable 5.2 at max effort, $4.64 a task, against $1.28 for Opus 5.5

Netflix’s write-up: netflixtechblog.com/fact-che…

Netflix’s AI fact-checker got cheaper and more accurate at the same time. The old setup ran one big reasoning model five times on every synopsis, each call with 100k+ tokens of context. When the reasoning was off, all five runs were off together. The new one uses two specialist judges. One asks “is this literally true?”, the other “does this give a new viewer a false picture?” Only the cases they’re unsure about get escalated. Cost per synopsis: down 52% False synopses caught: 77% → 86%

Coinbase CEO Brian Armstrong shared a personal AI Agent set up that lets him make decisions using his voice based on what his Chief of Staff Agent recommends. Here's how it works: 1/ One lead agent per area of focus, always on, each running sub-agents that write code, test it, chase reviews and coordinate with the team 2/ A chief of staff agent that oversees the leads 3/ One prioritised queue (a simple local web app) instead of dozens of separate chats, with screenshots and a video demo on every product review 4/ He clears it in batches, mostly by voice and keyboard shortcuts 5/ A “Brian brain” that learns his preferences, so first drafts need fewer corrections 6/ Every change still goes through the same human review and security checks I unpack this and more personal AI agent systems here: departmentofproduct.substack…