We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Agent Products Are Entering the Permission-and-Trust Test
•
3 min read
•
286 docs
Meta’s Muse launch puts agent access, approval, and failure handling under the microscope; Supermemory’s rapid shutdown adds a counterexample on retention and strategic fit. The brief pairs those signals with operating, stakeholder, and career tactics for PMs.
Coverage is incomplete: some monitored sources or documents could not be processed. This brief covers the available verified material.
Big Ideas
Agent products now need a permission architecture. Meta’s Muse is designed for task handoffs rather than chat: each user gets a dedicated cloud VM, Sentinel controls connected apps and accounts, and Muse asks before spending or sending messages. Internal testing nevertheless found private iCloud photos exposed, unreliable ticket and inventory monitoring, and mid-task logouts; the launch had already been delayed to address security. Meta’s “rule of two” avoids combining untrusted content, sensitive access, and external action, but still requires human approval gates. For any agent feature, specify what it can touch, who approves, when approval occurs, and what happens when it fails.
Standalone tools face a bundled-and-AI benchmark. A ProductManagement thread reports Miro acquired by Bending Spoons for $1.355B versus a last $17.5B valuation. In the same discussion, users cite moves to FigJam and AI-built prototypes, while others defend Miro’s differentiated interaction quality and report 250+ Miro users versus 10+ Figma users. The PM question is whether a product owns a workflow strongly enough to survive bundling—not whether its basic features can be copied.
Tactical Playbook
Productize repeatable AI work instead of automating judgment. Ramp’s creative queue saw more than 375 requests in six months. In a four-week overhaul, artists defined layouts, narrative patterns, and quality standards while builders encoded them; a Slack agent now returns editable decks, moving human work upstream to edge cases and new patterns. Apply the pattern to PM operations: choose a recurring artifact; encode its context, constraints, examples, and evaluation criteria; assign owners and feedback loops; route unfamiliar or defining work to experts. Judge success by whether output is grounded, editable, reviewable, and visible—not merely rendered. Ramp attributes a separate 48-hour launch to shared context and authority close to the work, not lower standards.
Use skip-levels for signal, not status. Bring what is working, friction, unclear requirements/priorities/ownership, risks leadership may not see, one improvement, and a concrete request for a decision or help; skip the PowerPoint. The payoff is longer-horizon context and less-filtered ground truth, not another status channel.
Case Studies & Lessons
Supermemory chose unshipping over launch reach. About one month after launching Company Brain, the team reported 500k+ X impressions, 2M+ across platforms, and hundreds of companies using it; it discontinued Company Brain and Nova and refunded charged users. Most of its first 100 signups churned. The founder cited worsening product clarity, doubt that Supermemory was best positioned to build the broader company-brain future, and conflict with infrastructure customers, then refocused on the memory API. Use retention, strategic fit, and channel conflict as continue/kill criteria; launch reach is not product proof.
Career Corner
Keep the reasoning path in the job. Shreyas Doshi warns that AI can increase answer and prototype volume while teams bypass the thought process and atrophy core product skills. Make the PM own the problem frame, evidence trail, and decision rationale even when AI supplies drafts. Aakash Gupta similarly argues influence is less commoditized than SQL, impact sizing, or technical fluency; his practical models are reciprocity, matching media richness to stakes, careful delivery, warm introductions, and attention to small credibility signals.
Tools & Resources
Agent-native launch pattern: cfo.ai. Users reply with a business, the agent does work in public, and the output becomes the next demo; the team also rebuilt the spreadsheet layer around agents rather than placing an agent atop legacy software. For agent products, design usage to generate inspectable proof and revisit the underlying workflow, not only the interface.
Agent Products Are Entering the Permission-and-Trust Test
Julie Zhuo
- When to pursue hyperpersonalization: Zhuo argues it is most valuable for high-frequency workflows where small frictions compound, especially across seams between tools, and for products where users want software to express their individual preferences. She places this shift on a spectrum from fixed use-case software to platforms and all-purpose builders, with increasing user freedom and imagination.
- Product principles: Make the interface malleable; let preferences go beyond predefined toggles through descriptive intent; let user feedback change the product; and treat cross-tool seams as part of the product experience.
- Dogfooding case: Zhuo found her remote workflow for managing eight AI-agent terminal windows cumbersome and six clicks deep, so she built a mostly working alternative in about 30 minutes and then tailored it with a custom grid, agent-status indicators, and support for both Claude and Codex. Because she was the target user, she could identify and remove friction and maintain a highly efficient feedback loop.
- Personalized learning case: For her children’s math and reading app, Zhuo used stories tied to each child’s interests, friends, hobbies, and recent trips, added lesson-level difficulty and content feedback, and introduced rewards based on their requests. She cites a study of 145 ninth-graders in which personalized algebra problems were solved faster and more accurately, with the largest gains among struggling students and benefits persisting after personalization was removed.
Product Management
- Valuation and growth signal: The thread reports Bending Spoons acquiring Miro for $1.355B versus a last valuation of $17.5B; the original poster frames the reset as a warning about heavily funded products pursuing hypergrowth that they cannot sustain.
- Standalone-tool economics are under pressure: Commenters report moving from Miro to Jira Whiteboard, FigJam, Confluence Whiteboards, or AI-built prototypes because alternatives are cheaper, “good enough,” or bundled with existing software; one commenter nevertheless reports 250+ Miro users versus 10+ Figma users at their company.
- AI is changing prototyping and collaboration workflows: A designer says their organization is using Figma less because of AI; another says Figma Make credits can be exhausted in a day, Claude is cheaper for similar work, and Make does not use existing design-system components without extra effort. Other comments say Claude-generated whiteboards, Claude Code, and MCPs are putting some visual-collaboration use cases at risk.
- Differentiation lesson for PMs: One Miro user cites fluid object movement and connector prediction as meaningful UX advantages, while another says basic whiteboarding is easier to copy; collaboration products therefore need to prove that distinctive interaction quality and workflow value justify standalone pricing over bundled or AI alternatives.
Hiten Shah
- Brand as software as an operating model: Hiten Shah argues that a guideline only records what a team has learned, while software can apply that knowledge to the next deck, landing page, or launch so work starts from accumulated company context. The proposed “institutional memory with hands” connects customer calls, product data, analytics, Slack, and Notion, carrying audience context, product truths, promises, approved examples, and escalation rules into the work.
- Implementation pattern: Treat the knowledge layer as a product with owners, builders, evaluation criteria, maintenance, feedback loops, and a roadmap; version, test, and continuously improve it. Automate proven, repeatable work, but route unfamiliar or defining decisions to human experts and keep review mandatory where reliability is uncertain; a partial output that forces a rebuild can be worse than no tool.
- Ramp case study: Ramp’s creative queue received more than 375 requests in its first six months, prompting a four-week overhaul focused on decks, whitepapers, and one-pagers. Artists defined layouts, narrative patterns, and quality standards while builders encoded them into a system. A Slack agent can turn an attached Markdown file, Google Doc, or Notion page into editable HTML, PDF, and PowerPoint outputs; the approach expanded to whitepapers, one-pagers, and landing pages, moving creative effort upstream toward new patterns and edge cases.
- Cross-functional launch lesson: A Ramp team spanning brand, product, engineering, communications, and partnerships moved from an opportunity to a named, designed, developed site with customer proof, assets, and a launch within 48 hours; the speed was attributed to shared context, authority, and craft being close together rather than to lowering standards.
Mind the Product
- Meta’s agent product bet: Meta launched Muse on September 8 as a personal AI agent available on the web, iOS, Android, and WhatsApp; unlike a chatbot, it is designed for users to hand off tasks such as drafting emails, booking flights, and making purchases. Each user receives a dedicated cloud virtual machine, while Sentinel controls what Muse may do; users choose connected apps and accounts, can interrupt the agent, and must approve spending or sending messages. Meta’s rollout sequence was model-first—Muse Spark in April, open-weight Muse Glimmer in August, then the consumer product.
- Agent safety framework and launch-readiness lesson: Meta’s “agents rule of two” says an agent should not simultaneously process untrusted content, access sensitive data or systems, and take an action that changes something or communicates externally; Meta describes this as only one layer that must be supplemented with human approval gates. Despite delaying Muse from April specifically to address security concerns, internal testing reportedly found safeguard bypasses that exposed private iCloud photos, unreliable ticket and inventory monitoring, and Meta CTO Andrew Bosworth being logged out mid-task. The PM takeaway is to evaluate not only whether an agent ships, but what it can access, who approves actions, when approval occurs, and how it behaves when it fails.
- Commercial and positioning choices: Muse offers 100 million tokens per week free for typical users, with $20-per-month and $100-per-month paid tiers; Meta says it plans to monetize primarily by taking a small cut of completed transactions rather than showing ads. Meta is positioning Muse for mass consumer adoption as a personal assistant, while Grokbot is positioned toward power users and enterprise work, illustrating different target segments and willingness-to-pay strategies. The broader market is also testing agent placement across browsers, standalone apps, and messaging, giving PMs competing distribution patterns to evaluate rather than a settled interaction model.
The community for ventures designed to scale rapidly | Read our rules before posting ❤️
- The smart-fridge concept bundles three difficult products—reliable inventory sensing, grocery purchasing, and health guidance. The recommended lowest-risk starting wedge is expiry detection plus meal suggestions using user-confirmed inventory; the team should prove repeat engagement with waste reduction before taking on custom hardware. Autonomous ordering makes a single detection error financially costly, while therapeutic diet guidance requires a higher trust bar.
- The proposed trust and autonomy rollout uses three stages: visibility and expiration reminders, a reviewable draft basket, and fully autonomous ordering constrained by budget and diet rules. The founder sets a 99% item-recognition target before scaling and treats an 80%-accurate order as a likely churn event because of misidentification risk.
- The business model deliberately treats hardware as a low-margin or subsidized distribution channel, with recurring premium nutrition subscriptions and transaction revenue from retail and delivery partners as the primary monetization.
- Early discovery feedback flags adoption and market-entry risks: one commenter rejects an in-home camera and additional AI in the home, another flags cybersecurity in Samsung’s comparable product, and a hardware commenter reports that distribution can overwhelm feature advantages.
Hiten Shah
- Supermemory discontinued its Company Brain and Nova products roughly one month after launch, despite 500k+ impressions on X, more than 2M impressions across platforms, and hundreds of companies using Company Brain; customers who had been charged were refunded. Early product feedback was also weak: most of the first 100 signups churned.
- The shutdown prioritized product clarity and strategic focus: adding applications had made it harder to explain what Supermemory was, the team judged itself poorly positioned to build the broader “company brain” future, and application-layer products risked making infrastructure customers view Supermemory as a competitor. Supermemory therefore refocused on its frontier memory API and infrastructure for agents.
- PM lesson: strong launch reach and initial usage do not outweigh poor retention, unclear positioning, or channel conflict; reassess whether a product strengthens the core strategy and customer trust, and be willing to unship it when focus and differentiation suffer.
The community for ventures designed to scale rapidly | Read our rules before posting ❤️
- Paid beta / client-funded R&D: The founder’s B2B SaaS has a working core but an incomplete roadmap; they proposed charging early customers $99/month against a $199 list price in exchange for regular feedback, arguing that payment is a stronger willingness-to-pay signal than free-beta feedback. A commenter described this approach as “client funded R&D,” saying upfront payment can validate the sale and motivate customers to try the product seriously.
- Conditional monetization rule: One response recommends charging from day one when users receive a service they will eventually pay for, but waiting until paid features are usable when the current users are not the revenue source. Another suggests keeping users free until the product works properly, then starting with a low price and increasing it over time.
- Founding-cohort mechanics: If offering a free or discounted pilot, one commenter recommends a contract with six months free and an extension if the product is not ready; a follow-up recommends explicit terms and a clawback if customers break the agreement after the six-month period.
- Free-tier design: A commenter says free users may help put a company logo on the product but often show less product value, while another gives a concrete example of a daily usage cap on free accounts versus hourly background checks on the paid plan.
Hiten Shah
- cfo.ai demonstrates an agent-native launch pattern: users reply publicly with a business or modeling task, the agent performs the work in public, and the output becomes the demo and proof of capability. Each subsequent user can provide another task, creating a compounding loop in which product usage generates evidence before others try the product.
- Product design implication: instead of placing an agent on top of legacy software, rebuild the core workflow—in this case, the spreadsheet layer—around how agents actually work.
Lenny Rachitsky
Lenny Rachitsky reported strong demand for in-person PM community after an event attended by more than 1,000 PMs; recordings of the talks are planned for his YouTube channel.
Shreyas Doshi
AI can help product teams produce more answers and prototypes, but bypassing the underlying thought process may atrophy essential product skills and invite the teams’ own obsolescence, according to Shreyas Doshi.
Product Management
- A PM with 12 years of experience across industries argues that product managers should not default to constant availability: use a “by appointment only” operating norm, reserve direct interruptions for the manager or manager’s manager, and let engineering counterparts handle incident response; the PM also argues that customer communication does not always need to come from product.
- Practical boundary-setting tactics include defining a genuine-emergency escalation path, turning off push notifications, and using a separate work phone that is physically put away after hours; one PM working across US, European, and Asian teams uses a timed lockbox and gives family the work number for emergencies.
- Availability should be calibrated to role and context rather than treated as universal: one proposed model is five-minute responsiveness during 9–5, up to two hours in the evening, no overnight expectation, and four-to-six-hour weekend response for simple follow-ups or production issues, with more flexibility needed around major launches.
Product Management
- Treat skip-levels as strategic context and relationship-building, not status reporting. They may serve to review or sense-check a manager’s performance and keep senior leaders connected to actual work, while giving PMs access to broader strategy, longer-horizon opportunities and risks, and context that may be filtered before reaching leadership. Use the time to explain decision rationale and trade-offs, discuss blockers beyond your manager’s ability to unblock, and ask what leadership considers most important next.
- Bring a concise, forward-looking agenda. Cover what is going well, unnecessary friction, unclear requirements, priorities, ownership or product direction, risks leadership may not see, a proposed improvement, and any concrete request for a decision, escalation, context, or blocker removal.
- Build trust before raising sensitive feedback, then use the relationship for growth. Early meetings can focus on understanding the director’s intent, listening, and building rapport; when raising problems, stay candid and professional, connect them to work or morale impact, and bring evidence plus improvement ideas rather than personal attacks. Skip-levels can also support mentoring, career direction, skill development, and building internal champions or sponsors.
Product Management
- Interview-prep tactic: A PM is building a tool that maps companies such as Google Maps and Spotify to provide industry context for product-sense questions, with Airbnb and Stripe planned; the approach helps candidates compensate for limited exposure beyond their own niche.
- Career-tool product pattern: Joey centralizes a candidate’s work history, matches it to job requirements, asks about missing information, and drafts role-specific resumes that can be checked against source evidence while learning the user’s writing preferences without inventing experience. Before sending resume text to an outside writing model, it strips direct identifiers; the product is invite-only while its builder gathers feedback.
Product Management
- In complex, heavily customized B2B products, UAT should validate customer-specific business processes, configurations, legacy data, and user flows in addition to the technical and functional coverage expected from QA/SIT. Unexpected behavior may require joint investigation because the customer holds context about business-process nuances, legacy configurations, and test data that the product team may not possess.
- To reduce surprises without treating every discovery as a preparation failure, introduce an earlier customer-facing QA or pre-UAT step that brings the customer into the process sooner for feedback before formal UAT.
- When each customer has a heavily customized legacy version and the product lacks documentation, test automation, code reviews, and CI/CD, UAT problems can reflect accumulated technical debt and unclear ownership rather than an individual PM’s preparation. Teams should distinguish preventable defects from customer-context discovery and address the underlying documentation, automation, and SME gaps.
The community for ventures designed to scale rapidly | Read our rules before posting ❤️
For product positioning and demo work, prioritize clarity over polish: consider paying for a clean deck or short product demo when it prevents confusion, iterate the narrative through real investor or customer questions, and defer a full rebrand or elaborate video until customers are pulling for it.
Hiten Shah
PM operating principle: standards scale only when teammates can apply and enforce them without the leader being present; otherwise execution remains dependent on that individual.
The community for ventures designed to scale rapidly | Read our rules before posting ❤️
- For an early B2B SaaS with a live core but an incomplete roadmap, charge once the product delivers clear value: payment provides a stronger willingness-to-pay signal than free usage, though charging can add friction for early adopters.
- Treat paid early access as a product-discovery loop: commenters argue that paying customers provide more candid feedback and force stricter prioritization; start with a small, reasonable price and iterate based on what customers say.
- One proposed implementation was $199/month list pricing with $99/month for the first customers in exchange for regular feedback, positioned between lower-cost tools and much more expensive enterprise offerings.
Hiten Shah
Hiten Shah’s local-AI experimentation pattern is to start with one ordinary machine assigned one job, repeat the experiment until the failure modes are understood, and then decide what additional setup is needed; he plans to demonstrate six local-model jobs on an ordinary laptop.
The community for ventures designed to scale rapidly | Read our rules before posting ❤️
- Use a 30-day focused B2B go-to-market learning loop when an early service business needs repeatable demand: choose one buyer type and one narrow service or pain, use direct outreach to learn objections, and test a small number of referral partners.
- Reduce adoption friction with a small, fixed-scope assessment instead of pitching the full managed service. Use prior client work as proof only when the buyer, problem, and scope are genuinely comparable, and frame the case study around the problem and outcome.
- Evaluate each experiment by qualified conversations, segment and contact reason, agreed next steps, and time spent qualifying—not by raw contact or reply volume. Add referral partnerships once the buying trigger and introduction process are clear.
Product Management
A product manager who moved from 10 years as a data engineer into media technology product management reports difficulty identifying new AI use cases despite weekly internal AI demos; they are considering targeted courses or hands-on PM-and-AI projects to build capability.
