Now Shipping: A Mind the Product podcast

AI’s biggest rivals are talking about slowing down. But will their release schedules tell the same story?

This week on Now Shipping, Mike Belsito breaks down Dario Amodei’s 3,800-word essay on why frontier AI development needs to slow down, the responses from Sam Altman and Elon Musk, and what it could mean for product teams.

If your roadmap depends on the next model being smarter, cheaper and just a few months away, how exposed are you? Mike explores the assumptions worth challenging, the questions to ask your AI vendors, and whether proposed safety standards could strengthen the biggest labs’ advantage.

Plus, Nina Olding, Staff Product Manager at CoreWeave and former Product Lead at Google DeepMind, shares her perspective on what PMs can control today: shipping AI responsibly through evals, transparency, human oversight and meaningful user choice.

Chapters
00:00 Watch the releases, not just the essays
00:48 The researcher’s warning that sparked the debate
04:22 Dario Amodei’s case for slowing down
06:01 Independent evaluators and a three-part safety plan
07:07 The criticisms: accountability, coordination and competition
08:23 Sam Altman and Elon Musk respond
09:46 The political reaction
10:32 What this means for product roadmaps and vendor selection
12:53 Will public promises change what companies ship?
13:45 Stress-testing your dependence on better AI models
15:04 Nina Olding: What responsible AI looks like for PMs
Featured links
Dario Amodei: We Must Pace the Frontier
Jacob Coxon’s resignation and AI safety warning — TechCrunch
Watch Jacob Coxon’s BBC interview
Sam Altman and Elon Musk back calls to slow AI development — The Guardian
AI CEOs say they need to slow development. But will they? — The Guardian
The original 2023 open letter: Pause Giant AI Experiments — Future of Life Institute

Does your roadmap depend on AI models getting better? Tell us what you’d rethink if progress slowed.

Your calendar, your inbox, your credit card. How much access would you give Meta’s AI agent?

This week on Now Shipping, Mike Belsito examines Meta’s pitch for Muse and the questions it raises for product teams building, buying or experimenting with AI agents.

Where should your product’s AI features live? What should an agent be allowed to do? And who approves its actions before a message gets sent or money gets spent?

Mike explores why permissions could become a reason users choose one AI product over another, and what product people can learn from competing approaches to putting AI into everyday workflows.

In this episode:
• Where AI belongs: chat panels, browser extensions or the apps people already use.
• Meta’s pitch for Sentinel as an approval gate for Muse’s actions.
• Why permissions and user control deserve a place in product comparisons.
• Four questions to ask about any agent: what can it access, who approves its actions, when is approval required, and what happens when it gets something wrong?

Would you give Muse access to your inbox and credit card? Tell us where you’d draw the line in the comments.

Now Shipping is Mind the Product’s weekly AI briefing for product people, hosted by Mike Belsito. Subscribe for more AI stories that matter to your work.

Oji and Ezinne Udezue are product leaders with more than 50 years of combined experience between them. Oji has held product roles at Twitter, Calendly, Atlassian, Typeform and Microsoft; Ezinne has led product at WP Engine, Procore and T-Mobile. Together they co-wrote Building Rocket Ships and now run ProductMind, helping companies work out what AI actually means for how they build. 

In this episode of Now Shipping, host Mike Belsito brings them in to unpack a story that kept repeating through the month: AI agents breaking out of the sandboxes built to contain them.

We discuss:
— How OpenAI's most advanced model went rogue during an internal security test, hacking into Hugging Face's production infrastructure without any human instruction, in order to cheat on a benchmark
— Why Moonshot AI's Kimi K3 didn't need to hack anything — it found an unlocked outbound connection, cloned the target repository from GitHub, and read the benchmark's solution straight off disk
— Why the "rogue intern" analogy applies to agents, and why the real failure sits with the people who build fragile boundaries, not the agents that find them
— The shift from zero trust to "negative trust" — treating every agent as having to reprove itself in the moment, rather than trusting it because it behaved well five minutes ago
— Why harness engineering, not prompt engineering or loop engineering, is where the next year of agentic product work will actually happen
— A five-part framework for product managers partnering with security teams: reach, reversibility, graceful exits, observability and provenance, and the kill switch
— Why an industry fixated on completion rate, rather than safe-stop rate and data provenance, risks repeating this month's incidents

This week we look into the launch of Slack Code. Salesforce's new product turns any Slack conversation into a dedicated workspace where AI coding agents such as Claude Code, Devin and GitHub Copilot can plan, write and ship code together, with a human still required to sign off before anything reaches production. 
Mike Belsito unpacks how the launch fits Slack's rough year, what it means that Anthropic was named a founding partner just three days later, and why product managers should be paying close attention to the permissions model behind every agent platform they adopt.
Key takeaways
Slack Code turns any Slack conversation into a dedicated workspace where coding agents can plan, write and ship code, splitting work into a conversation tab, a plan tab, a diff view and a live preview, with a human still required to sign off before anything reaches production.
The launch reads as counter-programming for a difficult year at Salesforce, whose stock has lagged the market and whose Slack CEO defected to OpenAI, prompting the company to go on the offensive with AI rather than retreat.
Vibe coding is no longer a fringe habit: it is already a $4.7 billion market projected to triple by 2027, with 63% of people doing it describing themselves as non-developers.
Slack's bet is that coding works better as a team sport than a solo terminal session, wagering that visibility and peer comment beat one person working alone with an agent in private.
The same openness that makes Slack Code powerful raises the stakes on quality: independent research puts the exploitable-vulnerability rate in AI-generated code at roughly 44%, even as AI-assisted developers ship three to four times faster than everyone else.
Barely three days after launch, Slack Code's neutral positioning shifted when Salesforce and Anthropic announced Claude Force, naming Anthropic a founding partner and making Claude the default model across the Slack and Salesforce stack.
For product managers, the workflow modelled by Slack's own VP of product, tagging in an engineer before asking for approval and taking one more pass on the agent's output, may become the template for contributing real code without pretending to be an engineer.
Chapters
(00:00) Introduction and the new one-story format 
(01:15) Slack Code launches 
(02:15) Inside the Devin demo 
(03:35) Why now for Slack 
(05:20) Vibe coding by the numbers 
(07:05) Multiplayer versus single-player coding 
(07:55) The AI slop problem 
(11:00) Permissions and sandboxing 
(12:05) The Claude Force twist 
(13:25) What it means for product managers 
(16:15) The model war question 
(16:30) Wrap-up
Referenced
Slack Code launch announcement, Salesforce (24 August 2026)
Cognition/Devin live demo with VentureBeat
Rob Seaman, EVP and GM of Slack
Katie Steigman, VP of Product, Slack
Jeff Wang, Cognition
Mario Rodriguez, GitHub Chief Product Officer
Veracode, 2026 security report on AI-generated code
Gartner, citizen developer forecast
Claude Force partnership announcement, Salesforce and Anthropic (26 August 2026)
Marc Benioff, CEO, Salesforce
Dario Amodei, CEO, Anthropic

Mike Belsito hosts Now Shipping, the weekly AI news show for product people, brought to you by the team at Mind the Product. This week's three stories cover a platform war breaking out over where the world's code lives, the first real fallout from AI content watermarking, and a set of behavioural numbers from Linear that quietly undercut the standard business case for AI adoption.

In this episode, we cover:
— Cursor launched Origin, a full GitHub competitor, on the same day GitHub suffered a six-hour worldwide outage with a 20% global failure rate.
—  Origin is deliberately interoperable with GitHub, which removes the switching cost that usually keeps frustrated teams on a platform.
—  Developer tooling is consolidating into a two-platform fight between Microsoft and SpaceX AI, which closed its $60bn acquisition of Cursor four days before the Origin launch.
—  Anthropic is now watermarking every response Claude generates, everywhere, in response to the EU AI Act's transparency code of practice.
—  The watermark uses SynthID Text, encoding a signature in low-stakes stylistic choices that survives light editing but not a full rewrite.
—  The backlash is loud but misdirected: other major model providers have signed the same code, so this is an infrastructure shift rather than a Claude-specific one.
—  Linear's first data report, drawn from over 127,000 paid users, shows AI adoption more than doubling in every function between January and June, with product managers climbing fastest.
—  AI now authors close to half of all issues created in Linear, up from fewer than one in a thousand two years ago.
—  Teams with a coding agent went from 21 pull requests a week to 65; teams without one moved from 8 to 10.
—  Planning time did not move at all, in any function, in a year when execution metrics rose everywhere else.
— Linear frames the result as a Jevons paradox: AI work landed as a new layer on top of existing work, and total time spent went up rather than down.

Chapters:

(00:00) Introduction 
(00:16) This week's stories 
(01:01) Cursor launches Origin 
(02:01) SpaceX AI's acquisition of Cursor 
(02:34) Interoperability and switching costs 
(03:04) A two-platform developer tools market 
(04:42) Anthropic starts watermarking Claude 
(05:15) What the EU AI Act requires 
(05:42) How the watermark works 
(06:35) The backlash 
(07:08) What it means for product builders 
(09:22) Linear's first data report 
(10:12) AI adoption across functions 
(10:37) AI-authored issues and pull requests 
(11:21) Planning time hasn't moved 
(12:30) Jevons paradox and rising workload 
(13:28) Wrap-up

Referenced:
Cursor: https://cursor.com
GitHub: https://github.com
Anthropic: https://www.anthropic.com
SynthID Text (Google DeepMind): https://deepmind.google/technologies/synthid/
EU AI Act: https://artificialintelligenceact.eu
Linear: https://linear.app
Linear's first data report
Mind the Product: https://www.mindtheproduct.com
Vote for Mike's SXSW session: https://tinyurl.com/belcetosxssw