Source: Lenny’s Summit 2026 talks, published on Lenny’s Podcast (YouTube, all fetched 2026-09-29):
- Claire Vo (ChatPRD, How I AI) —
raw/Everyone_s_shipping_more._Does_any_of_it_matter_Claire_Vo.md(youtube.com/watch?v=VM5kuvWgwDY) - Marty Cagan (SVPG) —
raw/Marty_Cagan_-_Strong_Opinions_loosely_held.md(youtube.com/watch?v=fF3lkTCM5-c) - Dan Shipper (Every) —
raw/How_to_run_your_product_team_like_a_research_lab_Dan_Shipper_Every.md(youtube.com/watch?v=DqF08Dz3nok) - Katie Dill (Stripe) —
raw/How_to_scale_intent_quality_and_artistry_with_Al_Katie_Dill_Stripe.md(youtube.com/watch?v=GLvFTMtw4Jk) - Robby Stein (Google Search) —
raw/What_it_takes_to_be_a_top_PM_today_Robby_Stein_Google_Search.md(youtube.com/watch?v=sTgM_sbLMNg) - Tamar Yehoshua (Atlassian CPO) —
raw/Roles_aren_t_converging_they_re_expanding_Tamar_Yehoshua_Atlassian_CPO.md(youtube.com/watch?v=BtK4kFI1LNo) - Elena Verna (Lovable) —
raw/The_rise_of_HI-ICs_Elena_Verna_Lovable.md(youtube.com/watch?v=fn8wnmpVqeI) - Karri Saarinen (Linear) —
raw/Why_I_took_the_summer_off_from_AI_and_what_I_learned_Karri_Saarinen_Linear.md(youtube.com/watch?v=Zn9NZ-r1-C4) - Anthropic panel: Ami Vora and Mike Krieger, moderated by Dan Shipper —
raw/Why_Claude_can_t_be_your_PM_yet_Anthropic_CPO_Panel.md(youtube.com/watch?v=sEXdyK6woKU)
Nine Lenny’s Summit 2026 talks on what product work becomes once building is no longer the constraint. The summit was held in September 2026 (Shipper mentions Fable 5.1 and GPT-6 Astra launching “last week”).^[inferred] The shared premise is that building is no longer the constraint: in Claire Vo’s words, “execution has outrun my ability to discover meaningful … commercializable products.” The talks split on what replaces the old discipline: new artifacts (convictions, evidence, labs pipelines), standards encoded into the tools, or deliberate rituals that keep people learning. The one direct disagreement, about whether roadmaps survive, is set out side by side below. Most figures are conference claims about the speakers’ own companies. Ramp’s talk from the same summit has its own article: The Software Factory — Warp and Ramp.
Key Takeaways
- The scarce thing is judgment about what to build. Vo has “40 Grok bots” and PRs “up 3x” but is “out of good ideas”; Robby Stein calls decision-making the PM’s core craft; Katie Dill warns that “AI makes things feel finished really really fast, all too often prematurely.”
- Filtering moved to after the build. Dill: the quality filter that used to run through every stage “now needs to happen post build when it is a lot harder to say no”, so someone must act as an editor who uses the product as a user would. Ami Vora: it is now “faster to just build three versions and try them out.”
- Separate exploring from executing. Dan Shipper’s labs team expects to throw away about 90% of what it makes; the product team expects to adopt about 10% of it. Anthropic keeps parked projects running in an eval harness so a model jump shows up on its own.
- Put standards into the machinery. Dill: “the object of design is no longer the screen. It is the system itself.” Stripe replaced a design-docs MCP with a design-system CLI after “three different people could put in the same prompt and get three different results.”
- Automate the routine, protect the learning. Karri Saarinen: build software factories if you like, but as an organization “don’t become a software factory”, because “the output is not the product.” Linear’s answer is weekly rituals where everyone finds and fixes a defect and critiques new features in the open.
- Org charts follow the cost of building. Elena Verna: “When the cost of building falls below the cost of coordinating, the org chart should change.” Tamar Yehoshua: whether a PM codes (“rowing”) or directs (“steering”) depends on the product’s phase. Mike Krieger: the PM is still needed because “I don’t think Claude is a convener yet.”
- Train on a schedule. Atlassian’s quarterly AI builder weeks have run 1,000+ people through and left 120+ workflows in use.
The roadmap disagreement: Vo versus Cagan
Both are opinion. Cagan answers Vo by name (“despite Claire’s optimism”).
| Claire Vo — “Build your last road map” | Marty Cagan — “I guarantee your road maps are not going away” | |
|---|---|---|
| Core claim | Under “roadmap zero” every visible feature is “plausible and buildable”, so effort stops being a useful way to prioritise. “An AI factory plus an old road map will get you to those three traps at machine speed.” | The question was never whether roadmaps exist but what they are used for. His “single biggest regret” is underestimating “how powerful and deeply rooted the desire for predictability was.” |
| What goes wrong | Three traps: backlog (AI builds every item, which is not progress), parity (every competitor reaches the same obvious product), churn (ship, see noise, abandon, never compound). | Roadmaps and PRDs enable “thinking that we know more than we really do”; used as requirements for engineers, that is the “project model” behind failed products. Cheaper engineering does not fix it: “the tokens aren’t free.” |
| What to do instead | Convictions about where the market is going; evidence defined up front (“what would I need to see to stop?”); a factory to reach reality fast; allocation of tokens and people. Label each ship a probe, experiment or promise. “Durable convictions but disposable features.” | Do discovery first. Once evidence shows what will achieve the outcome, the PRD or roadmap is “just your communication device”, harmless and useful. Separate “building to learn” from “building to earn” (a distinction he credits to Jeff Patton). |
| Does a roadmap survive? | A different one: “I do still think you need a road map, but … more about ambition.” And “I do not mean build your last plan.” | Yes, as a communication device after validation. |
Where they overlap: both reject a list of features with guessed impact and fixed dates that the team treats as a promise. The live difference is whether the artifact keeps the name and form of a roadmap.
Talk by talk
Claire Vo — “roadmap zero”
- Confession. Built a semantic “product graph” for ChatPRD in almost one shot, feature-matched competitors, then kept it feature-flagged off because she lacked conviction: “code is abundant customer trust is not.” She re-architected ChatPRD “70 times because I could” and stopped tracking issues.
- Why roadmaps worked before. Engineering scarcity filtered bad ideas: the bottom of the list, “below the cut line”, never shipped.
- Good versus bad stubborn. Good: stay with the problem, revise the solution. Bad: keep shipping and move the goalposts because tokens are cheap.
- Next year is “the ambition game”. Her suggested OKR for an AI transformation: how many big swings the team takes each month, “with the presumption that most of them won’t work out.”
Marty Cagan — ten regrets (only the AI-relevant ones)
- Two kinds of product organisation were on stage all day: the “project model” and the “product model”.
- He understated business viability in the first edition of Inspired, and says it matters more for AI products.
- The most important “why” is “why are people not using our product?”, which he credits Stein for raising.
- “In many companies, process is used as a substitute for thinking”; he worries LLMs will be used the same way.
Dan Shipper — run product like a research lab
- Labs versus product. A labs team explores new model capabilities and runs many parallel experiments; with AI it can be one person. The product team improves and scales what already works.
- “Two-slice team”. One or two people: a “pirate” (“slop cannons who are absolutely obsessed with finding value”) paired with an “architect” who turns a messy prototype into an extensible system.
- Make the discarded 90% pay. Publish what you tried as content, and use experiments to feed an early-adopter programme.
- Pipeline and promotion. Every’s pipeline tracker in Notion is reviewed weekly at all-hands. Promotion criteria: people use it and come back; it is still “10x better” a month later; it is affordable at scale.
- Worked example: “Kate bench”. Shipper gave Fable three years of the editor-in-chief’s copy edits; the result became a “Kate pass” on Every’s company agent. After an architect rebuilt it with a dashboard, “she did 12% less work on these types of edits than she did the month before.”
- Precedents he cites. Claude Code, MCP, skills and Claude Design came out of Anthropic Labs; Codex was built by a small team outside the main app, launched as a desktop app in February 2026, and was later merged into ChatGPT.
- Rule number one: “never make any major life decisions within 30 days of … your first encounter with a frontier model.”
Katie Dill (Stripe) — scale intent, not just consistency
- Three watch points: LLMs return the most probable (past, popular) answer; the “temptation of done”; and work so cheap it is treated as disposable. Unchecked, they produce “zombie UI”.
- Four recommendations: have a point of view (Stripe’s is optimism); encode standards into the machine; “refuse to confuse done with good”; unleash creativity and “protect the strange”.
- Stripe’s design-system CLI. It “makes the AI far more obedient”, loads documentation “at the right time and place to avoid context rot”, and ships full templates and flows rather than components. “The old system scaled consistency, but the new system needs to scale intent.”
- Quality bar. Asked “What’s the quality standard for something made with AI?”, her answer is the same bar as anything else; one design crit produced 17 fixes. The event’s opening animation took 56 AI-assisted iterations.
- Better inputs, stressed outputs. Put brand beliefs and source material in the prompt, push past the first result, and use adversarial agents to critique.
Robby Stein (Google Search)
- Understand people first. Jobs-to-be-done interviews (Christensen’s Competing Against Luck), and ask non-users why they are not using the product.
- Stack-rank opted-in feedback with a model. Google keeps a library of feedback from users who opted in. The top theme in one shopping case was “you should have asked me about my son and the height” for a backpack query; building clarifying questions became “one of the biggest things we’ve done in terms of engagement.”
- An agent that uses the product. Built with Google’s Antigravity, it asks questions as a user, screenshots the answers and grades them against a rubric (correct LaTeX for maths, a visual tray for visual topics); the team now has agents that find and increasingly fix breakages.
- What he wants to try next: extracting jobs-to-be-done from interview transcripts, and an agent that interviews users with the same method.
Tamar Yehoshua (Atlassian) — roles are expanding
| Project | PM role | What changed | Reported result |
|---|---|---|---|
| Confluence Remix and Slides | Rowing | A PM who had never coded merged 26 PRs in a month using an engineer-built harness in a clean, isolated repo; PM-led evals; Figma MCP plus a coding agent fixed design mismatches; test creation went from half a day to 10 minutes | 2x eval throughput; about 14 design bugs fixed an hour; Remix in 6 weeks and Slides in 8, against about 6 months before AI |
| RovoClaw (new product) | Rowing, then steering | A PM and a designer vibe-coded a working alpha; once engineers joined, the PM stopped coding to set direction and unblock; an agent writes the weekly updates | Moved “much faster” after the switch |
| Jira (large existing codebase) | Steering | No PM code in production; a Loom recording creates work items that start a cloud coding agent; Slack feedback triaged by an agent and sent to a coding agent; 900+ user-study items triaged by a Rovo agent | About 3x throughput; 22 user-facing features in about 10 weeks |
- AI fluency index. Six capabilities (such as using the tools, automating data insights, prototyping, technical literacy), each scored from 1 “curious” through 3 “capable” to 5 “pioneering”. Everyone is expected to reach level 3 over time; it is a development tool, not a promotion ladder.
- AI builder weeks. Once a quarter: external speakers, internal peers teaching, and a project, one theme each time (prototyping, evals, building agents, checking in code). “Builder week in a box” is on Atlassian’s site.
- Measure what ships. PRs deployed to production, not written; features delivered to customers. “Everything that you read on X, don’t believe.”
Elena Verna (Lovable) — the high-impact IC
- A high-impact IC finds the problem, decides, executes across functions, ships and owns the outcome. She treats AI as “average intelligence” in every specialty, so one or two areas of real craft are enough.
- Five preconditions: information flows freely (each Lovable team has a Slack agent with named human “parents”); authority travels with accountability, including room to fail (her failed experiments cost Lovable “many millions of dollars”); scope is not defined by function; pay and status are decoupled from headcount; and people pick manager or IC, not both.
- Her survey: 42 of 51 people-managers wanted to go back to an IC role. “Management needs to be a career path, not a promotion.”
Karri Saarinen (Linear)
- He stopped following AI news for the summer and found on return that “nothing has changed much” in what it takes to build a great product.
- Automate what you don’t learn from. Linear’s agent investigates errors through Datadog and Sentry and proposes a fix for an engineer to check. A “watcher” agent sends him a daily digest of what customers say about their AI workflows.
- Rituals that train the eye. “Quality Wednesday”: every week each person spends time in the product looking for a quality defect to fix and share. “Feature roast”: an optional meeting anyone can join, hosted by the team building a new feature, where people “critique the whole feature” as nitpicky as they like.
- “At Linear we don’t run experiments”; people are told to use their intuition, which he says is learned from working on things and listening to customers. His summary: “context becomes the product.”
Anthropic panel (Ami Vora, Mike Krieger)
- The PM role survived contact. Krieger moved to an IC role early this year; a PM lead insisted his project needed a PM, and a week after one joined he told her she was “100% right”.
- Claude runs internal software. Anthropic runs on Claude Tag, and “so much of the software we interact with is like built maintained and iterated on by claude”, including a status UI for a four-workstream project.
- Overlap is allowed early. Vora prefers “five overlapping products” to one over-constrained one, consolidating once something fits the market. Krieger: each labs bet has a “bet lead” who decides to double down or wind down. The panel describes a foundations team unifying memory, MCP and file storage, which Chat and Cowork did not share a few months earlier.
- Agent-native products. “Everything a human could do an agent should be able to do”, built on shared primitives rather than bolted on; a side panel is an acceptable first step.
- Park, don’t kill. Anthropic’s 2024 computer-use prototype stayed in an eval harness, and the team saw a model leap (captions: “37”) when it started “succeeding more often than it’s not”. Shipper’s term for the opposite failure: “capability blind.”
Try It
- Tag your next five shipped items as probe, experiment or promise, and write down for each the evidence that would make you stop (Vo).
- Name a “labs team of one” who tries each new model within a week and reports to the product team at a weekly review (Shipper).
- Keep one abandoned idea running as an eval and re-run it on every new model (Krieger).
- Give every shared team agent a named human owner (Verna’s “parents”).
- Run one quarterly builder week with a single theme and a project at the end (Yehoshua).
- Start a weekly find-and-fix-one-defect ritual, shared in a team meeting (Saarinen).
- If agents get inconsistent results from your design docs over MCP, test a CLI that serves templates and flows instead (Dill).
- Marketing: have a model stack-rank a batch of customer comments or survey answers and act on the top theme, as Google did with opted-in feedback.
Open Questions
- Most figures (Atlassian’s 26 PRs and 3x throughput, Every’s 12%, Lovable’s “many millions”) are self-reported conference claims without methodology.
- Vo and Cagan offer opinion, not data; neither cites outcomes from teams that dropped or kept roadmaps.
- “37” in the Anthropic panel is presumably Claude 3.7; the captions do not confirm it.
- Speaker attribution in the Anthropic panel’s auto-captions is uncertain for some answers.
Related
- The Software Factory — Warp and Ramp — Geoff Charles’s talk from the same summit
- The New Shape of Product Work (Andrew Ambrosino) — “implementation is cheap, taste is the bottleneck”
- Dianne Penn — Evals Are the New PRDs — the eval-first counterpart to Krieger’s parked projects
- Running an AI-native engineering org (Fiona Fung) — Anthropic’s own account of the shifted bottleneck
- Dan Shipper — The AI Paradox — Shipper’s earlier Lenny’s episode
- CLI vs MCP — where Stripe’s design-system CLI fits
- Figma MCP Server — the design-to-code loop Atlassian used
- Browser and Computer Use in Practice (Claire Vo) — agents that use the product as QA