We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Personal agents make product craft matter again
@sriramk sees "a come back of product management" in how personal agents are designed. Instinct, Dot, Muse, and Grok bot each involve real product work to connect the UX with the model underneath and with the services it can reach . He says he "rarely care[s] about the underlying model" but cares a lot about small touches, like Muse's activity notification or Instinct's thumbs-up emoji to acknowledge a message. He contrasts this with the "upgrade to latest model" path the industry is otherwise on . Lenny Rachitsky shared it with the comment "PM ascendant" .
The strategy question is what happens to companies these agents route around. In "The Amazon Trap," @mvernal (an initial investor in Instinct and on its board) argues that Amazon spent 15 years training shoppers to start on Amazon instead of Google, and its $80B ad business is the reward . Now it is reacting badly to Instinct and Muse and leaning on an "Alexa for Shopping" tab . His case: horizontal agents win because they hold all of a user's context. That leaves Amazon choosing between giving up ad revenue and annoying its customers, while Walmart and Shopify can take share by supporting agents . His proposed way through is a "Prime+" tier at about $299 a year with agentic purchasing only available there. It would give loyal customers a way in and replace lost ad revenue without betting the company . Scott Belsky agreed: "you can't fight customer preferences for horizontal agents, you must innovate to meet them" .
Broad versus specialized agents
An a16z conversation pushed back on the all-in-one model. One speaker connects his agent to everything so it can join context across chat history, GitHub, Salesforce, and his calendar . The counterpoint: the more work you hand a general agent, the more understanding you give up, and "agents don't have any responsibility" . That speaker expects vertically focused sub-agents with quality checks, possibly coordinated by a chief-of-staff agent . Replit spent nearly a year making its product deployable in customers' own clouds because of data-leak worries. Speakers agreed that making agents useful at work is still largely unsolved . OpenRouter added that it is surprised how little benchmarking enterprises do, and it expects internal AI teams to build more evals .
On the build side, Teresa Torres found that when an agent edits a tree, comparing the input and output trees can produce several valid change sets, and only one shows what actually happened. Her fix is to give the agent rules for meaningful moves and log every move as it builds .
Keep the judgment yourself
Aakash Gupta disagrees with the complaint that Claude Code dulls you. If 90% of your job is pressing enter, "the model is making your decisions" . He cites an MIT Media Lab EEG study: people who wrote on their own first and then brought in the model showed the highest engagement . His real workload is deciding which of a dozen prototypes "deserves to exist" and saying no, because "when building is free, shipping everything buries the three things that matter" .
A practitioner post shows the same shift landing on teams. A PM leading two PMs and three designers says that after adopting GitHub and Claude, designers were expected to deliver close to production-ready pages, covering every error state, breakpoint, and accessibility case, without training for it. Meanwhile engineering keeps setting and changing direction .
Leadership and growth
- The Standard-Setter. John Cutler argues that pushing people harder produces evidence "that people can respond to pressure," not evidence of impact. Many people respond better to clarity, trust, autonomy, and good feedback . The style looks good to executives, while the costs (exhaustion, churn) land on the team .
- Upskilling for today. Shreyas Doshi says mid-career PMs wrongly see upskilling only as preparation for the next job. Building stronger leadership skills would free up time in their current role, which breaks the "too busy to upskill" loop .
- Flat teams. Belsky praised "talent dense, aligned, flat, and empowered teams" . The example he quoted was Nikita Bier dropping his assigned work to fight spam. Bier himself notes the model is "very hard to make … work in a larger company" .
- Occasional expertise. Hiten Shah: you don't review pricing or plan a launch every day, but competitors keep moving. AI changes what it costs to keep that expertise current .
- Stripe’s acquisition of OpenRouter preserved OpenRouter’s autonomy over its brand, roadmap and product; the deal was also presented as a way to speed execution, strengthen go-to-market and create product synergies.
- OpenRouter’s product thesis is that AI products need to differentiate beyond prompts by combining multiple models with unique data and services, while reducing vendor lock-in; a marketplace can help customers find models suited to their tasks and increase pressure to lower costs.
- Enterprises are exploring models beyond proprietary frontier offerings for cost and differentiation, while building internal AI practices; the speakers expect greater use of benchmarks, evaluations and cost-per-task research, noting that benchmarking remains less common than they expected.
- Replit spent nearly a year working on deployment in customers’ own clouds/on-prem as concerns about data sovereignty, security and agent-related data leakage grew; the speakers said making AI genuinely useful and productive at work remains unresolved.
- The discussion surfaced a trade-off in agent design: broad agents can connect context across chat, GitHub, Salesforce and calendars, but may leave users with less understanding and diffuse responsibility. One proposed direction was vertically focused agents with quality checks; enterprise designs also need access isolation and safer agent-to-agent communication.
- Speakers presented task-specific decision models with structured outputs as more controllable than general models; Replit’s example was a cost estimator that classifies prompts into price buckets, and they argued that proprietary-data classifiers could avoid some model churn. OpenRouter and Cognition also launched fusion tools/models, described as combining model families to broaden idea search and reduce costs; OpenRouter’s initial application focused on deep research.
Personal-agent design is a renewed product-management opportunity: the view Lenny Rachitsky amplified is that products such as Instinct, Dot, Muse, and Grok bot need product work to connect user experience with model capabilities and access to services . The post highlights interaction details—such as Muse activity notifications and Instinct’s thumbs-up acknowledgment—and contrasts them with simply upgrading to the latest model; its author says they rarely care about the underlying model but do care about those details .
- Validate demand before finishing or launching a product: research the market early and solve a real need rather than expecting customers to be persuaded they need the product.
- Use small experiments to test behavior, not just interest—for example, put up a landing page and payment link, or aim to get one stranger to pay even $10.
- Follow customer feedback beyond acquisition: compare what sales promised with the onboarding experience, and learn from customer calls and lost-deal reviews in a setting where sales, product, and support communicate.
Shreyas Doshi argues that mid-career product people should treat upskilling as a way to improve their current work-life balance, not only as preparation for a future promotion or job change: stronger product leadership skills can help them handle expanded scope in less time, while postponing development because they are busy perpetuates the cycle of being too busy to upskill and too busy because they have not upskilled.
When an agent generates tree edits, comparing only the input and output trees can yield multiple change sets, even though only one captures the meaningful change; a semantic merge, for example, is different from a deletion followed by source reassignment. Teresa Torres’s proposed fix is to define rules for meaningful moves and log each move as the agent builds the tree.
One PM leading two PMs and three designers reports that engineering frequently sets and changes direction, leaving decision ownership unclear even though the PM is expected to own product decisions. In an account-management redesign using GitHub and Claude, expectations shifted from design work to AI-assisted pages close to production-ready, covering requirements, error states, responsive breakpoints, and accessibility; the designers lacked training for that technical responsibility. The post raises the need to define product/design/engineering responsibilities and account for training, validation, and learning time in delivery expectations.
- Aakash’s AI-assisted discovery workflow pairs Perplexity market sizing with three expert calls and NotebookLM synthesis, then consolidates support calls, app reviews, sales calls, and interviews into an AI customer-intelligence system. Use that context to write a PRD and build a prototype for customer testing before engineering commits a sprint; then synthesize prototype feedback from transcripts, recordings, and notes using Teresa Torres’ discovery method. He also recommends using an agent to monitor competitor launches, customer sentiment, pricing, and feature gaps, with a weekly rhythm of reviewing intelligence, shipping and monitoring, synthesizing learnings, and generating insights.
- Aakash argues that PMs should delegate execution to AI without handing over product judgment: he points to choosing which prototypes deserve to exist and saying no to low-priority work as the decisions that remain central when building becomes easy.
- A product owner who had delayed making an explainer because conventional options seemed expensive and complicated tried Claude; the described workflow used UI elements created in Figma, imported as an SVG, and simpler prompts, which worked better than specific instructions.
- A commenter judged the result’s motion and cuts insufficiently polished for high-profile public-facing marketing, while potentially good enough for internal use; the right bar depends on the use case.
- Pressure-based “high standards” may prompt people to stretch in some contexts, but that response alone is not evidence of impact; people vary in what motivates them, with clarity, trust, autonomy, meaning, mastery, useful feedback, and worthwhile effort among alternatives to pressure.
- A Standard-Setter’s visible ambition and team wins can conceal costs borne by teams, including exhaustion, resentment, churn, and the effort of managing the leader; the approach depends on positional power until the leader must demonstrate their own ability or encounters someone they cannot bully.
- In a discussion about moving from video-game production into product management, a producer with seven years of experience described transferable work that included recommending alternatives based on resource calculations, pushing prioritization and tradeoffs, owning schedules and flagging deadline risks, maintaining vision alignment, running retrospectives and coaching leads, and coordinating vendors; they had worked with teams of 50–100.
- One PM’s advice for aspiring PMs was to build product thinking, problem-solving, analytical, stakeholder-management, and strategic skills, and demonstrate them through portfolio deliverables tailored to the person’s goals and gaps.
Hiten Shah argues that product teams need to keep expertise in areas such as pricing and launch planning current even when they use it infrequently, since competitors keep moving; AI changes the cost of maintaining that expertise.
An automotive systems engineer moved into product management at a battery-swapping startup because experience with batteries, chargers, and IoT was more valuable for that role than having a PM title; the author learned PM responsibilities on the job. After a layoff, the author moved toward Growth Marketing by recognizing overlap between PM work and marketing—positioning, go-to-market, understanding customer funnels, prioritization, and cross-functional execution. The career lesson: consider adjacent roles and make transferable skills visible; sideways applications can reveal a viable path to an internal move, rather than treating a title or sector mismatch as proof that those skills do not transfer.
Scott Belsky endorses the view that companies cannot fight customer preferences for horizontal agents and must innovate to meet them. The underlying product strategy argument is that a single personal agent drawing on context across email, calendars, files and messages can serve many needs better than separate task-specific agents; Amazon risks losing customers or market share if it resists that shift. As a possible transition, the article proposes testing agentic shopping in a premium Amazon tier, giving loyal customers agent access while creating a paid route to preserve value as shopping behavior changes.
A former investment banker seeking a permanent PM role says they built an EdTech website using Cursor and Claude Code, partnered with a software-engineer co-founder, and launched a university pilot; they also have two contract roles training and improving AI models for finance.
Paul Graham highlighted the prediction that claims AI models cannot conjecture, theorize, or explain are already outdated and may soon describe no current models—a caution for product teams making plans around those capability limits, not evidence of a specific product change.
Scott Belsky praises talent-dense, aligned, flat, empowered teams . The example he cites is an organization where people can drop their assigned work to tackle an urgent problem, raise the alarm, and solve it or rally others—an approach described as difficult to sustain in larger companies .