🔵 Claude can Dream and Coinbase says the future is one person product teams

Plus: Linear’s design engineer on how to build agents with taste, Atlassian’s SaaSpocalypse narrative violation and more

Hi product people 👋,

This week, Claude released a new feature called Dreams. We’ll dig into that along with new releases from OpenAI and Google Docs.

Plus, after announcing 14% job cuts, Coinbase’s CEO says that the future is “one product teams,” where the roles of product, design, and engineering all merge together under an “AI pod” that manages a fleet of agents. But is this a realistic vision of the future? We’ll take a look at his plans in more detail to try to find out.

Have a great Friday and enjoy the weekend!

Rich

Watch on YouTube | Follow on Substack Notes

---

Key reads and resources for product teams

New from the Department of Product Substack this week:

Deep Dive - Spotify’s new Natural Language Ad API and other Natural Language Interfaces Explored
Spotify just let advertisers create campaigns through the API with natural language. Now 30+ companies - from Canva to Uber to Shopify - are rethinking entire workflows around natural language interfaces. This latest Deep Dive Product explores how NLI is transforming product design, which parts of your product might be ready for it, and why the future of UI might just be letting AI handle the clicks while humans focus on judgment calls.

New in the AI Library

• Identify opportunities for ChatGPT Apps for your product - This prompt builds a structured idea generator that takes your brand, customer types, and business model as inputs, then produces 8-10 specific ChatGPT app concepts tailored to your business

• Evaluate and remove AI feature bloat - use this prompt to design a customer feedback synthesis that identifies and removes feature bloat.

Interview - Uber’s CEO on Agentic Coding and why product people are making changes to Uber’s codebase
“Now, they’re just going in and they are vibe coding, and an engineer is going to review the code, but essentially the product person is going directly into the code base, so to speak, or going directly with an agent into the code base.” Uber’s CEO on replacing policies with outcomes, why agentic coding is changing who writes the code, and the bet that there won’t be one physical AI model to rule them all. (Decoder Podcast)

Resource - A library of sounds you can use on your product’s components to bring it to life
Designed by Raphael Salaja, this is a rich library of component sounds that simplifies sound management in web applications to help bring your product to life. (WebKit Audio)

Design - Linear’s Design Engineer on How to build agents with Taste
Emil Kowalski shares how he taught Claude to understand his design taste by packaging animation rules into a “skill file” - easing curves, scale values, timing guidelines, and all. (Emil Kowalski)

Strategy - Emotional churn explained
Emotional churn, when users mentally check out before officially leaving, can be the silent killer of your B2B business. It happens weeks or months before actual cancellation, hiding in apathy and radio silence, argues Jake Stein. (Substack)

Skills - A guide to Agent-Native Product Management
AI is now doing the drudgework of product management - turning three-hour analytics investigations into minutes of conversation with Claude. In a practical guide for Every, Marcus Moretti shares the agent-native techniques he uses to run product management for Spiral, including two installable skills (strategy and product-pulse) that let PMs focus on the actually interesting parts of the job. (Every)

Resource - WebUI Comparison Benchmark
A comparison of how different models handle the same UI prompts, side by side. Useful if you’re wondering which model to use for different parts of your product design process. (WebUI)

Analysis - Coinbase’s CEO says the future is “one person teams”
Coinbase is laying off 14% of its workforce, but CEO Brian Armstrong frames it as necessary preparation for an AI-powered future. The company is restructuring around smaller “AI-native pods,” eliminating pure management roles, and flattening hierarchy to move faster through market cycles. The future, he says, is “one person teams”. But is he right?

---

New product features and innovation this week

Anthropic released three new capabilities for Claude Managed Agents this week that help autonomous systems work more effectively without constant human oversight.

“Dreaming” is a scheduled review process that analyzes past agent sessions and memory to identify patterns (things like recurring mistakes, converged workflows, team preferences). It restructures memory to keep it high-signal as agents accumulate experience, enabling self-improvement across sessions. You control whether dreaming updates memory automatically or requires your approval first.

You can think of it as the agent’s being awake during the day (sessions) where it tries to write memories as it goes. And at night, as it sleeps, it compresses, deduplicates, removes stale or contradictory info, and surfaces cross-session insights that weren’t obvious from a single interaction. Just as a human might do when they dream.

Anthropic says companies including the likes of Harvey and Netflix are already using this. Harvey uses dreaming so agents carry knowledge between sessions, things like file format workarounds and tool quirks. The result was roughly a 6x improvement in agent completion rates.

Outcomes define success criteria as a rubric. A separate grader evaluates the agent’s output against your standards in an isolated context window, then surfaces specific gaps if corrections are needed. The agent iterates until the work meets your bar. This works for both objective requirements (file formats, completeness) and subjective criteria (brand voice, design guidelines). Internal testing showed 10-point improvements in task success, with +8.4% gains on Word documents and +10.1% on PowerPoint files.

An example of multi-agent orchestration analysing ecommerce checkout flows

Multiagent orchestration lets a lead agent break complex work into parallel tasks and delegate each to specialist agents with custom models, prompts, and tools. Specialists share a filesystem and context, enabling the lead agent to check progress mid-workflow. Full tracing in the Claude Console shows which agent did what and why.

ChatGPT gets a new plugin for Sheets and new Voice models, pitching voice as the “interface between people and products”

OpenAI released ChatGPT for Google Sheets, an add-on that embeds natural language AI directly into spreadsheets via a sidebar panel.Instead of manually building formulas or structures, you describe what you need in plain English - “create a budget tracker” or “analyze spending by category” - and ChatGPT builds, edits, and updates the sheet accordingly. It already has over 100,000 downloads with a 5 star rating so far, but it’s still early days.

This week, OpenAI also released three new audio models designed for developers building voice-enabled applications: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper.

OpenAI is explicitly pitching the new Realtime-2 model to product teams, describing it as an “interface between people and products”. They say that as voice becomes a more natural way to use software, three distinct patterns are emerging on how developers are using voice APIs:

1. Voice to action - where people describe what they need and the assistant uses tools on behalf of the user to complete the task

2. Systems-to-voice - where software turns context into live guidance e.g. “Your inbound flight is delayed, but you can still make your connection. I found the new gate, mapped the fastest route through the terminal, and your bag is still expected to transfer”

3. Voice-to-voice, where AI can help live conversations continue across languages, tasks, or changing context

Some product teams, they say, are combining all three approaches to build an end-to-end voice powered UX. You can watch some of those use cases in action here.

Other updates

Google Docs now lets users set persistent custom instructions for Gemini that automatically apply to all future interactions. Instead of repeating preferences like “use bullet points” or “keep it professional” in every prompt, users input these rules once in the Gemini side panel. Gemini then remembers and follows these instructions across all conversations within that document.

Spotify has released a Command Line Interface tool called “Save to Spotify”. It lets you connect tools like Claude Code generate podcasts from any text source and then store it to your library. Spotify’s CPO Gustav Söderström said:

The use cases feel obvious once you have it: for example a morning briefing built from your calendar and inbox. A deep dive on your class notes before an exam. A travel itinerary narrated for your flight.

Is this the end of the swipe?

Bumble may be ditching the swipe as a user interface for dating. They posted a cryptic “Dear swiping, it’s over” message on Instagram, sparking speculation that the dating app might ditch the swipe. Bumble reported a 21% drop in paying users this week and the company has hinted at a major overhaul called “Bumble 2.0,” including AI-powered features and in-person events. But will ditching swipes actually fix the app’s problems?

If it works, we could see other product teams in other categories follow suit and actively add friction to their core UX to encourage a more mindful experience. Which, given the state of the average person’s doom-scrolling induced, dopamine-riddled brain, may be a good thing.⁠

---

Tools you can use

Cognitive Bias Index - A tool for identifying and mitigating the systematic errors in thinking that affect the decisions and judgments that humans make.

Marcus - AI prompt library and thinking system for leaders. Apply concepts from over 100 business books in your prompts and stress-test important leadership decisions.

Kanwas - a centralized hub where product teams accumulate, connect, and build on product context.

---

📈 Product data and trends to stay informed

The SaaSpocalyse narrative was violated on a couple of fronts this week. Atlassian revealed that revenues grew 32% to $1.8 billion in its latest quarter, accelerating from 23% growth the previous quarter. Their CEO Mike Cannon-Brookes also revealed that customers using their AI Assistant Rovo generate twice the annual revenue compared to non-users.

And despite fears over the disruptive effects of AI, Figma was named as one of the fastest growing companies in Ramp’s latest AI tools report. This only uses data from Ramp customers but assuming Ramp’s customer profile is reasonably AI native, it indicates that those companies are not ditching Figma in favor of other AI tools.

Pinterest now has 631 million MAUs globally, up 11% year-on-year - tenth consecutive quarter of double-digit user growth.

AI now writes most of the new code at several major new tech companies and new data suggests that the hype surrounding Claude Mythos was justified. Anthropic says With the help of Claude Mythos Preview, the Firefox team fixed more security bugs in April than in the past 15 months combined.

Amazon says almost 60% of users now use autocomplete when searching - and the company’s product teams are working on a new hybrid search which will push users to Rufus if it detects questions more suited to a conversation.

---

Paid subscribers get the full DoP Substack including: The Knowledge Series for sharpening your tech / AI skills, the AI Prompt and Skills library and DoP Deep dive reports for in-depth analysis to learn lessons from the world’s top tech companies.