Why Claude Can’t Be Your PM (Yet) Anthropic CPO Panel

Why Claude Can’t Be Your PM (Yet) | Anthropic CPO Panel ![rw-book-cover](https://d34adp677peecb.cloudfront.net/static/images/article2.74d541386bbf.png) 🔗 **Source:** [Why Claude Can’t Be Your PM (Yet) | Anthropic CPO Panel](https://www.youtube.com/watch?v=sEXdyK6woKU) > [!note]+ Document Notes > tags: [product-management, agent-native-software, ai-product-strategy, team-adaptability, parallel-experimentation] > author: Ami Vora, Mike Krieger, Dan Shipper > source: Lenny's Podcast (YouTube) > reference: https://www.youtube.com/watch?v=sEXdyK6woKU > date created: 2026-09-29 10:32 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Why Claude can’t be your PM (yet) | Anthropic CPO Panel >  ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--why-claude-cant-be-your-pm-yet--anthropic-cpo-panel) ## 📝 Summary - **What happened.** Lenny's Podcast published a 28:48 recording of a live panel from the Lenny and Friends Summit (San Francisco, September 10, 2026), uploaded 2026-09-29. Per the video description, Anthropic's Ami Vora and Mike Krieger answer questions from moderator Dan Shipper. The auto-transcript has no speaker labels, so this analysis attributes a line to a named panelist only when the content makes the speaker clear, and says "a panelist" otherwise. - **Commercial interest, stated up front.** Both guests speak for Anthropic and discuss Anthropic's own products and internal practice: Claude, Claude Tag (the transcript renders it "cloud tag"), a chat and Cowork product pair, and a model the transcript calls "Fable." The panel's central claim, that product managers stay necessary while Anthropic's models improve, comes from people whose employer sells those models and whose hiring plans depend on the answer. Nothing in the transcript discloses a sponsor of the panel itself. - **The claim.** Faster building raises the value of the PM job instead of removing it. A panelist calls the role a "convenor" that connects people, customers and safeguards. The skills that lose value are interaction-level design taste and up-front prediction. The skills that gain value are judgment under ambiguity, relentlessness, adaptability and clear decision ownership. - **The product claim.** Software should be agent-native, meaning anything a human can do in it an agent can also do, and interfaces that the agent can reshape. Ami and Mike (as best the transcript shows) describe an internal project view that Claude built and the team edited with Claude. - **The portfolio claim.** Run overlapping experiments ("five overlapping products" beats one over-constrained product), fund a foundations team for shared memory and primitives, park projects the model can't yet do, and re-run them against each new model with an eval harness. - **Why it matters.** It's a first-party account of how one frontier lab staffs product work, and it gives testable practices: parked-project evals, shared primitives, named decision owners. - **Risks and tradeoffs.** Evidence is anecdotal (one project, one PM arrival, one Instagram usage figure). Overlapping products risk user confusion, which the panel acknowledges at 00:24:58 without a solution. Speaker roles are not verified by the transcript. - **Bottom line.** The convenor and judgment claims are plausible and consistent with the anecdotes given. None is measured. Treat the panel as a lab's stated operating philosophy, not as evidence about PM demand across the industry. ## 📖 Main content **The opening question.** Dan Shipper frames the mood swing: a year ago the talk was that AI would make PMs obsolete, and now the talk is that everyone must become a PM. Both guests answer that the job hasn't gone away, but the way it's done has changed. **The bridge argument.** One panelist defines product as a bridge between human problems and available technology. Technology now changes about every two months, where it used to change every five or ten years. Human problems change slowly. The practical rule that follows is to stay obsessed with the human problem and treat what you know about the technology as disposable. **The convenor argument.** Mike describes moving to an individual-contributor building role in early 2026 and pushing back when a PM lead asked for a PM on a near-ship project. Once the PM joined, Mike says, the "glue and connective tissue" held: enterprise contingencies, customer success enablement, safeguards, and keeping builders heads-down. A panelist adds that Claude catches cross-organization overlaps (the transcript mentions finding one "via the slash search") but is not yet a convenor that schedules meetings on its own initiative. **What lost value.** A panelist says the craft of predicting where a thumb lands on a phone screen and where a button belongs is no longer worth the review time, because building three versions and trying them is faster. The panelist ties this to identity: a skill you're proud of can become the thing you must let go, the way the innovator's dilemma applies to companies. **What gained value.** Adaptability, judgment in the face of ambiguity, relentlessness (because forking paths multiply when building is cheap), and framing the chaos so people engage with it. The panel adds a structural counterweight: exploration inside a team needs a named decision owner, called a "bet lead" in Anthropic Labs, who says double down or wind down. **Agent-native software.** Mike lays out a progression: an AI sidebar, then AI-powered features, then agent-native architecture where everything a human can do an agent can also do, then interfaces the agent can reshape. The transcript names a write-up on agent-native architectures by an author whose name the captions garble ("every literally wrote the book on it"); the description doesn't identify the author, so treat the attribution as unresolved. Mike says the team turned that write-up into a skill and used it to guide the work. For a company with an existing SaaS product, the advice is to build shared primitives so agents and the REST API run through the same plumbing, then evolve gradually (a side pane first, a malleable personal landing page later). **Exploring versus stabilizing.** Ami describes a shift from a career valuing simplicity, reliability and predictability to encouraging "shots on goal." Teams need real conviction; assigning an unexcited team to a direction produces a weak product. A foundations team solved shared memory, MCP and file storage across surfaces so experiments feel complementary. Leaders should say out loud that three bets on one problem can feel demoralizing. **Parking projects.** Mike describes the first computer-use product built in 2024 as too weak to help either automation or education, so Labs parked it and re-ran it on each new model in an eval harness. The transcript says a "37" model showed a jump when the harness transcripts showed success more often than failure. The lesson given: build early, externalize the result as an eval, and avoid capability blindness. **Coherence.** Asked how to fold a winning experiment back into one product, neither guest claims a solved answer. Mike cites Instagram's strong power law in tab usage (main feed near 80%), and says half the job is saying no to another sidebar item until something is proven. **Contested or unverifiable in this source.** The Instagram 80% figure has no source in the transcript. The "37" computer-use jump has no eval numbers. The claim that the PM's arrival prevented dropped items rests on one anecdote. The models named "Fable" and "Astra" appear only in audience-polling banter, and the transcript gives no capability detail for either. ## 🗒️ Notes #### ⏱️ [00:00:06](https://www.youtube.com/watch?v=sEXdyK6woKU&t=6) Opening frame: obsolete PMs, or everyone a PM - Moderator Dan Shipper sets up the whiplash between two narratives about PMs. - Asks what a great PM at Anthropic looks like and whether the role is changing. > [!quote] [00:00:06] > a year ago we uh were told that AI might make PMs obsolete #### ⏱️ [00:00:48](https://www.youtube.com/watch?v=sEXdyK6woKU&t=48) Product as a bridge, and a technology cycle of two months - Technology used to change every 5 to 10 years and now changes about every 2 months (a speaker's estimate, not a measurement). - The human side changes slowly, so the plan is to obsess over the problem and stay fresh-eyed on the technology. > [!quote] [00:00:48] > a bridge between the real problems that people have in the world and the technology that you can use to solve it #### ⏱️ [00:02:13](https://www.youtube.com/watch?v=sEXdyK6woKU&t=133) Roles blur, and Mike's "do we need a PM" story - A panelist compares the moment to early product days: no definitions, so a generalist solves problems until hitting a wall, then uses scaffolding. - Mike moved to an IC role early in the year. A PM lead insisted the near-ship project needed a PM. A week after the PM joined, Mike conceded the point. - The captions render the concession name as "cat"; the speaker most likely means Ami. Unverified. > [!quote] [00:02:56] > do we need a PM? Like, Claude's got it. #### ⏱️ [00:03:39](https://www.youtube.com/watch?v=sEXdyK6woKU&t=219) What the PM carried - Items listed: serving prosumers through large enterprises, enabling customer success for real-time questions, safeguards, keeping people on track. - Builders stay in "full-on cloud mode" (Claude mode) while someone else makes sure everything connects. - Speed raises the operational bar for the role. > [!quote] [00:04:22] > this hat is increasingly important #### ⏱️ [00:04:22](https://www.youtube.com/watch?v=sEXdyK6woKU&t=262) "Why didn't you just have Fable do it?" and the convenor - Shipper asks what the human does that a model couldn't. Answer: Claude already supplies connective tissue and catches cross-org overlap, but is not yet a convenor. - A panelist gives a joke scenario, a Claude scheduling a meeting for two people, and says the capability exists and the setting is simply off. > [!quote] [00:05:05] > somebody still needs to bring the clouds and the people together to like get the work done #### ⏱️ [00:05:46](https://www.youtube.com/watch?v=sEXdyK6woKU&t=346) Judgment and relentlessness with cheap building - Tools expand what's possible to build, so choosing what to build from limited information matters more. - More forking paths means more need to make sure the chosen thing ships. - Each model release is described as a new landscape, neither utopia nor disaster. #### ⏱️ [00:07:11](https://www.youtube.com/watch?v=sEXdyK6woKU&t=431) Skills that stopped mattering - Ami spent years learning to model a user's situation and button placement because building was expensive. Now three versions get built and tried. - Identity cost: pride in an old skill delays adopting a new method. > [!quote] [00:07:54] > I don't think anyone's ever going to ask me that question again. #### ⏱️ [00:09:19](https://www.youtube.com/watch?v=sEXdyK6woKU&t=559) Traits: tolerance for change, judgment, relentlessness - Guest says they've never worked at a model company and re-tests their own usefulness every few months. - Talking openly about adapting reduces the sense that one person carries it alone. > [!quote] [00:09:19] > raising your ceiling on tolerance for change and being willing to embrace change is super important #### ⏱️ [00:10:45](https://www.youtube.com/watch?v=sEXdyK6woKU&t=645) Framing the chaos and naming a decision owner - Mapping the exact future is a control move that locks a person out of trying new things. - Emotions around the change are acknowledged as part of the work. - Anthropic Labs calls the accountable role a "bet lead," who says double down or wind down. > [!quote] [00:10:45] > frame the chaos so it feels safe and it feels plausible to engage with it #### ⏱️ [00:12:09](https://www.youtube.com/watch?v=sEXdyK6woKU&t=729) From human-only to agent-native software - Shipper asks what goes in a product versus an agent. - Mike's timeline: sidebar assistant, AI-powered features, agent-native (everything a human can do, an agent can do), malleable interfaces. - Mike says they fed an agent-native write-up into a skill during a year of related work. The author's name is garbled in the captions. > [!quote] [00:13:35] > the interfaces actually become something that is malleable by the agent too #### ⏱️ [00:14:17](https://www.youtube.com/watch?v=sEXdyK6woKU&t=857) A project view built by Claude, and provenance - A project with at least four independent work streams (the captions say "orc streams") has a Claude-built UI that a TPM and then the whole team use. - Anyone can change the display along with Claude instead of waiting on an internal or third-party team. - Open issue raised: who can update the data, and how to keep provenance. > [!quote] [00:14:17] > if you didn't like that particular display, you're not bound to like however it was determined to be built #### ⏱️ [00:15:41](https://www.youtube.com/watch?v=sEXdyK6woKU&t=941) Advice for a SaaS company: build shared primitives - Probe a new product by asking an agent to do something a human could do and watch whether it works through shared plumbing. - Evolve gradually: side pane, then a malleable landing page on the same primitives. - Products feel "bolted on" when the agent doesn't share the product's infrastructure. > [!quote] [00:16:25] > you can tell when things feel super bolted on #### ⏱️ [00:17:08](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1028) Serving the whole spectrum while exploring - Shipper asks how to push the frontier while enterprise customers need stability. - The answer is to meet people where they are, explore first, then make winners coherent for people who want stable products. - The panelist encouraged predictability in earlier roles and now trades it for exploration. > [!quote] [00:17:51] > It's not like stuff is stable and we can just build the right answer. #### ⏱️ [00:18:34](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1114) Shots on goal, conviction, and shared foundations - The failure state is over-constraining a product based on a theory before it meets the market. - A team assigned a direction it isn't excited about yields a weak product. - A foundations team unified memory, MCP and file storage so experiments on top feel complementary. - Saying out loud that three bets can feel demoralizing is treated as part of the leadership job. > [!quote] [00:18:34] > take a lot of shots on goal > [!quote] [00:20:41] > you should be able to get your memory wherever you are within our products #### ⏱️ [00:22:08](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1328) Parking projects and re-running them per model - Product-market-fit signals stay basic: do people use it, like it, return, talk about it, get value. - The 2024 computer-use product was parked and retried on each new model. - Labs counts "we showed the model isn't ready" as a win if it becomes an eval. > [!quote] [00:22:50] > we actually just had it like running in an eval harness #### ⏱️ [00:24:15](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1455) Audience poll on model versions, and folding experiments back - A show of hands asks who has used "Fable" (many) and "Astra" (fewer). The moderator says people often judge an old model version. - Merging a winning experiment (tab, separate app) has no accepted answer. - A panelist says cognitive load shifts to users when products offer tools and ask them to choose. > [!quote] [00:24:15] > Have you used the latest model? > [!quote] [00:25:40] > it's easy to just transfer a bunch of cognitive load to the user #### ⏱️ [00:26:22](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1582) Instagram's power law and saying no - Mike cites strong power-law tab usage at Instagram, with main feed around 80% and Explore moving from 10 to maybe 15%. - Half the job is refusing new sidebar items or tabs until something is proven, while not killing it early. > [!quote] [00:26:22] > half of the job is saying no to another thing in the sidebar or another tab or another thing #### ⏱️ [00:27:04](https://www.youtube.com/watch?v=sEXdyK6woKU&t=1624) Next year: more builders, and a closed capability gap - Guest one hopes more people can build, answer their own questions and start businesses, with small teams and solo empowerment. - Guest two wants to close the gap between what models can do and what most people use them for. - Success is defined as bringing non-engineers along in a way that "doesn't feel disempowering." > [!quote] [00:27:48] > the gap between what the models are capable of and like what most folks are using them for > [!quote] [00:28:30] > do long horizon work but not in a way that feels disempowering ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. - [Why Claude can’t be your PM (yet) | Anthropic CPO Panel](https://www.youtube.com/watch?v=sEXdyK6woKU): the source video, 28:48, uploaded 2026-09-29 by Lenny's Podcast, 1,328 views at capture. - Lenny and Friends Summit, recorded September 10, 2026, San Francisco. Named in the description; no URL supplied there. - Ami Vora and Mike Krieger (Anthropic) and Dan Shipper (moderator). Names come from the video description. - [Anthropic](https://www.anthropic.com): employer of both guests and vendor of the products discussed (Claude, Claude Tag, chat and Cowork surfaces). - [Model Context Protocol](https://modelcontextprotocol.io): the "MCP" mentioned at 00:20:41 as one of the divergent systems the foundations team unified. - Innovator's dilemma (Clayton Christensen): named at 00:08:36 as an analogy for personal skill lock-in. No link is given in the transcript. - The agent-native architecture write-up mentioned at 00:12:52: the author's name is garbled in the captions and the description doesn't identify it. Find the original before citing. - Model names as transcribed ("Fable," "Astra," "37," "Sonnet 45," "Fable 7"): confirm against Anthropic's model documentation before reuse; the transcript gives no versions or dates. ## 🗣️ Paraphrase The panel argues that product management gets harder, not obsolete, when models can build in days. Technology now shifts about every two months while people's problems shift slowly, so the PM's job is to hold the human problem steady and throw away what they learned about the last technology. A PM who joined Mike's near-ship project caught the enterprise, customer-success and safeguard work that builders in full-on Claude mode would have dropped. The panel doesn't think Claude can convene people yet, though it already connects cross-team overlaps. What no longer pays is fine-grained interaction design done in review meetings. Building three versions and trying them is quicker. What pays now is judgment with thin information, relentlessness across forking paths, adaptability, and a named decision owner inside each bet. On product shape, the panel wants agent-native software: anything a human can do, an agent can do, on shared primitives, with interfaces that agents and users can both reshape. Anthropic's own project views are partly built and edited by Claude. For portfolio management the panel prefers overlapping experiments over premature consolidation, backed by a foundations team for shared memory. Failed experiments should become evals that get re-run on each new model. Merging a winner into one coherent product remains unsolved. Mike's Instagram experience suggests the main tab takes most use, so most of the job is saying no to extra tabs. ## 📰 Newsletter **Your next teammate might be a convenor.** At the Lenny and Friends Summit, two Anthropic product leaders told a room of PMs that faster building makes the coordination job more important, not less. Their example is one project where a PM arrived and caught the enterprise, support and safeguard work that the builders would have dropped. The takeaway for WordPress teams: when contributors can ship a patch or a block in an afternoon, the scarce work is release coordination, docs, translation handoffs and support enablement. Two practices are worth stealing. First, park the experiment the model can't do yet and turn it into an eval, so you can re-run it on the next release instead of forgetting it. Second, build shared primitives so an agent and the REST API use the same plumbing, which is the WordPress Abilities API argument in different words. Worth noting: the guests work for the company selling the models, so read the "PMs stay essential" line as a stated philosophy, not a labor-market finding. Links: [the video](https://www.youtube.com/watch?v=sEXdyK6woKU), [Model Context Protocol](https://modelcontextprotocol.io), [Anthropic](https://www.anthropic.com). ## ✅ Action items #### DevRel - Developer advocacy: run a talk on "parked projects as evals," with one worked example that shows a failed prototype re-run against a new model release. - Education: add an office-hours session on the convenor role, using the panel's PM-arrival story as the case. - Developer experience: audit one docs or SDK surface for "bolted on" agent access. Check whether an agent can do every task a human can through the same API. - Developer marketing: don't cite the Instagram 80% main-feed figure or "every two months" as fact in community material. Both are speaker estimates without sources. - Community management: name the bet lead or decision owner in any program that runs overlapping experiments, and say aloud that overlap is deliberate. - Internal/external/cross-project: ask Anthropic contacts which agent-native architecture write-up the panel referenced, and get the real title. #### OSPO - Policy and governance: write the rule for who can update project data when both humans and agents can. The panel raised provenance at 00:14:59 without answering it. - License compliance: no direct action from this source. - Security management: agents that can do everything a human can do widen the permission surface. Review least-privilege scopes for any agent-native tooling. - Software inventory: record which internal views or dashboards are agent-generated, and who maintains them, so ownership doesn't vanish. - Developer support: publish guidance on re-testing old "the model can't do X" conclusions against the current model, with a dated log. - Community engagement: no direct action from this source. - Strategic alignment: track "experiments retired to eval" as a portfolio KPI next to "experiments shipped." #### Community - Maintainers: assign one convenor per release cycle for cross-team glue (docs, translation, support), separate from the release lead. - Contributors: give contributors a one-page "what happens after merge" checklist, since the panel's lesson is that dropped connective work is the failure mode. - Sponsors: when sponsoring parallel experiments, ask for the retirement criteria and the eval that will be kept. - Ecosystem: no direct action from this source; nearest relevance is the shared-primitives argument for plugin and host interoperability. #### WordPress - Core: no direct action from this source; nearest relevance is the "agent can do what a human can do" parity test for the admin and REST surfaces. - Core Performance: no direct action from this source. - Core AI: use the parity test as an Abilities API review criterion: does every admin action an editor can take have an agent-reachable ability? - **Core Program Team:** name a convenor for cross-team dependencies in each release, and record which items they caught that would otherwise have dropped. - Mobile: no direct action from this source. - Playground: a Playground blueprint that re-runs a parked prototype against a newer model would demonstrate the eval-harness idea. - Community: meetup talk on "the job after builders get fast," with the convenor and bet-lead roles as discussion prompts. - Training: add a lesson on re-testing stale conclusions about model capability, using the 2024 computer-use example. - Marketing: no direct action from this source. - Support: prepare support enablement earlier in the release cycle, since the panel's example lists customer-success readiness as dropped work. - Design: note that interaction-detail reviews are the skill the guest says lost value. Test whether building three variants beats a review meeting for one small pattern. - Documentation: document agent access next to human access for each admin task, so parity gaps show up in docs review. - Photos: no direct action from this source. - TV: no direct action from this source. - Polyglots: no direct action from this source; nearest relevance is that translation is part of the connective work a convenor tracks. - Accessibility: agent-generated interfaces need accessibility checks at generation time. Add that gate before any malleable-UI pilot. - Meta: no direct action from this source. - CLI: no direct action from this source; nearest relevance is that a CLI is a natural shared primitive for agents and humans. - Hosting: no direct action from this source. - Tide: no direct action from this source. - Openverse: no direct action from this source. - Themes: no direct action from this source. - Plugins: no direct action from this source; nearest relevance is plugin authors exposing abilities so agents use the same path as the UI. #### The WPOCC - Programs: no direct action from this source. If The WP Open Community Collective funds a convenor stipend, define the role by the connective tasks in the panel's example. - Projects: no direct action from this source. - Funding experiments: fund overlapping small experiments with a stated retirement rule and a kept eval, per the panel's shots-on-goal approach. ### FAIR - Technical independence: agent-native tooling that depends on one vendor's hosted model creates a single point of dependency. Document it. - Policies and covenants: require provenance for data agents update, which the panel flagged but didn't resolve. - Productized governance artifacts: add a "decision owner" field to any charter or scorecard that covers parallel experiments. ## 🧪 Hypotheses 1. Software projects that add a dedicated coordinator for cross-team readiness (support, docs, safeguards) ship with fewer post-launch escalations than comparable projects without one. 2. Building three prototypes and testing them reaches a usable interaction design in less calendar time than one design review cycle, for UI decisions that reverse cheaply. 3. Products whose agent and REST API share one primitive layer show fewer agent-versus-human capability gaps than products that bolt on an agent feature. 4. Prototypes shelved with a stored eval get revived and shipped more often within 12 months than prototypes shelved without one. 5. Running overlapping experiments with a shared foundations layer produces fewer user-reported inconsistencies than overlapping experiments with separate memory and file systems. 6. In a mature product, one main surface keeps a majority share of usage after new tabs launch, as the Instagram anecdote implies. ## 📊 Methodology - **H1.** Retrospective study of 30 projects from one organization's tracker. Code each as coordinator present or absent, then compare post-launch escalations per 1,000 users, controlling for team size and launch scope. - **H2.** Within-team experiment. For 10 UI decisions, randomly assign a design review or a three-prototype test. Measure days to a decision and a 5-user usability pass rate. - **H3.** Audit 20 SaaS products. For 15 fixed tasks, record whether an agent completes the task through the same path as the UI. Compare products with a documented shared API layer against those without. - **H4.** Cohort comparison from an internal experiment registry. Tag shelved projects by presence of a stored eval, then track revival and shipment at 6 and 12 months. - **H5.** Survey plus support-ticket coding across two product lines, one with shared memory and one without. Count "feels disconnected" tickets per 1,000 active users. - **H6.** Pull per-surface usage shares from analytics before and after each of the last 5 tab launches. Test whether the main surface share stays above 50%. - **Controls.** Pre-register thresholds, use the same rater for coding, and report failed replications. ## ❓ Questions - Who was the speaker at each turn, and which claims belong to Ami Vora versus Mike Krieger? The captions don't say. - What did the PM actually prevent from dropping in the anecdote at 00:02:56, and how would anyone verify it? - Does the "convenor" role need a PM title, or can a builder with strong coordination habits fill it? - What evidence supports the 2-month technology cycle and the Instagram 80% share? - How does a team fold an experiment back into one product without breaking users' muscle memory, which the panel admits is unsolved at 00:24:58? - Who owns data provenance when both people and Claude update a project view, as raised at 00:14:59? - What is the eval format behind the parked computer-use project, and could outside teams reuse it? - How does the panel's "five overlapping products" stance square with enterprise buyers who need stability? ## 🔮 Future research - **PM demand under agent tooling.** Longitudinal hiring and role data would test whether the convenor claim holds beyond one lab. - **Malleable interfaces.** No source in the panel tests whether user-edited, agent-edited UIs stay usable, accessible and secure over months. - **Eval reuse.** A public registry of parked-project evals would let teams re-run stale "can't do" conclusions against new models. - **Consolidation patterns.** The panel names tabs versus separate apps as options, and neither has a documented success rate. - **Speaker-level analysis.** A speaker-labeled transcript would let researchers separate each guest's claims from the moderator's framing. ## 💡 Implications - **Practice.** Assign a named convenor and a named decision owner per bet, and log what the convenor catches. Re-test old capability conclusions on each model release. - **Policy.** Data updated by agents needs provenance rules and a named human accountable for the schema. The panel raised the question and left it open. - **Security.** Agent parity (an agent can do anything a human can) enlarges the blast radius of a compromised agent. Apply least privilege per ability. - **Education.** Teach judgment under ambiguity and coordination as explicit skills. The guests say interaction-detail craft matters less, so curricula that center it need a rebalance. - **Funding.** Fund foundations work (shared memory, shared primitives) separately from frontier experiments, as the panel describes doing internally. - **Vendor literacy.** A first-party panel about the vendor's own products is a good exercise in separating stated philosophy from evidence. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level on the transcript body (6,659 words, 292 sentences by a punctuation count, heuristic syllables): about grade 8.7. Flesch Reading Ease about 73, which is plain conversational English. Spoken, unpunctuated captions inflate sentence length, so the real grade is likely lower. #### 👀 Visual learners - Present: the description promises a live panel, so viewers likely see the speakers and any slides. The transcript carries no description of visuals. - Improve: add a one-slide diagram of the four-step progression (sidebar, AI features, agent-native, malleable) and a chart of the Instagram tab shares. #### 🔊 Aural learners - Present: the whole panel is spoken conversation with audience laughter and live banter. - Improve: a 2-minute audio summary that lists the practices (parked-project evals, shared primitives, bet lead) would help listeners who can't rewatch. #### 🗣️ Verbal learners - Present: analogies (bridge, innovator's dilemma, ocean and horizon) and concrete stories (PM arrival, parked computer-use product). - Improve: define "DRI," "TPM," "eval harness," "MCP" and "primitives" in one clause each. #### 🧩 Kinesthetic learners - Present: one applicable test is described: ask an agent to do a human task in a product. - Improve: give a 3-step exercise, for example "pick five tasks, try each through an agent, log the gaps." #### 🧠 Logical learners - Present: an explicit progression and a clear before/after claim on the role. - Improve: supply the data behind the 2-month cycle, the 80% share and the "37" jump so the argument can be checked. #### 🧑‍🤝‍🧑 Social learners - Present: live audience polling ("how many have used Fable?") and a moderator-led exchange. - Improve: add discussion prompts in the description, such as "who is the convenor on your team?" #### 🧘‍♀️ Solitary learners - Present: short enough to watch alone; each answer stands on its own. - Improve: add self-check questions like "which of your skills is a button-placement skill?" #### 🌿 Naturalistic learners - Present: the ocean and horizon image at 00:06:29 is the only nature reference. - Improve: nothing here links to environmental or sustainability angles. Energy cost of running many parallel agent experiments is an unaddressed topic. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies page landmarks and player controls, plus a description that states the event, date and guests. - Improve: the visuals are unknown from the transcript. If slides or on-screen text exist, they need a text equivalent under WCAG 2.2 SC 1.1.1. #### Keyboard navigation - Present: the YouTube player is keyboard operable with visible focus, so SC 2.1.1 and 2.4.7 pass through the platform. - Improve: nothing in the video adds barriers. #### Contrast and readability - Present: unverifiable from a transcript. Live stage recordings usually show speakers, not text. - Improve: any lower-third names or slides should meet SC 1.4.3 (4.5:1 for text). Captions at the platform level are user-adjustable. #### Text resizing and reflow - Present: the player scales and captions resize independently. - Improve: burned-in text, if any, fails SC 1.4.4 and 1.4.10 because it can't resize or reflow. #### Media and captions - Present: auto-generated captions exist, and a live panel ran with audience cheering and laughter marked as [laughter]. - Improve: the captions garble proper nouns: "Enthropic" for Anthropic, "cloud" and "clawed" for Claude, "cloud tag" for Claude Tag, "Ammy" for Ami, and "cat" where a name was likely spoken. That fails the intent of SC 1.2.2 (Captions, Prerecorded) because names carry the meaning. The transcript has no speaker labels, so listeners can't tell who is talking, which also affects SC 1.2.2 and 1.3.1 intent. SC 1.2.5 (Audio Description) matters only if meaningful visuals exist. Correct the captions and add speaker names. #### Forms and interactions - Present: none in the video. - Improve: not applicable in-video. The audience poll at 00:24:15 is a hand raise with no accessible equivalent for remote viewers. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (caption accuracy on proper nouns), **1.3.1** (no speaker identification in the transcript), **1.1.1 / 1.2.5** (possible visual content with no text or audio equivalent, unverified). The description lists no chapters ("CHAPTERS: none"), which weakens navigation by topic, an aim of SC 2.4.5 applied by analogy to timed media. Corrected captions with speaker labels are the highest-impact fix. ## 🌍 DEIB assessment #### Cultural breadth - Present: the panel covers customers from prosumers to large enterprises and cites a global product audience only in passing. - Improve: every example is US-tech-centric (Instagram, Photoshop, San Francisco). Add examples of teams in other regions or in non-software organizations adopting agents. #### Inclusive language - Present: plain language and no exclusionary terms. "Democratizing" at 00:27:48 frames access as a goal. - Improve: "clawpilled" (00:27:48) and "back when the dinosaurs were in the earth" (00:01:30) are insider or age-coded idioms. Replace with a plain description. #### Neurodiversity support - Present: a moderated Q and A with one question per turn is predictable and easy to follow. - Improve: no chapters. The topic changes at 00:12:09, 00:17:08 and 00:22:08 could be chapter markers so viewers can jump. #### Representation - Present: two guests and a moderator, with one guest describing a colleague's insistence on a PM. - Improve: the panel is about roles that affect many groups, but no voice from a customer, a non-engineer builder, or a contributor outside Anthropic appears. The stated goal of bringing "a person doing research in parallel" along is spoken about, not represented. #### Psychological safety - Present: guests normalize uncertainty, name emotions around change (00:10:45), and say demoralizing overlap is acceptable to discuss openly (00:21:24). - Improve: the identity cost of obsolete skills is raised without a support path. Pair it with concrete retraining resources. ## ✅ Logical validity #### Strong reasoning - **Causal with mechanism, 00:02:56 to 00:04:22.** The PM's arrival is tied to specific catches: enterprise contingency, customer-success enablement, safeguards. Naming tasks makes the claim checkable in principle. To strengthen: log the actual items caught. - **Analogical, 00:08:36.** The innovator's dilemma maps cleanly from firms to individuals: skill strength breeds reluctance to change. To strengthen: say what evidence separates strength-driven reluctance from ordinary risk aversion. - **Practical inference, 00:22:50.** Retrying a failed capability against each new model in a fixed eval harness is a sound way to detect a step change, and the "37" episode is a concrete instance. To strengthen: publish the pass rate before and after. - **Self-limiting statements, 00:24:58.** The guest says the team hasn't fully figured out folding experiments into one product. This concession raises credibility of the surrounding claims. - **Structural argument, 00:11:26.** Chaos in exploration paired with one accountable owner is a coherent org-design claim: it separates who explores from who decides. #### Fallacies - **Appeal to anecdote, 00:02:56.** "do we need a PM? Like, Claude's got it." followed by the PM's later value. One project, one PM, one before-and-after. Why weak: no comparison group, and the speaker who first doubted the need has a stake in the conclusion. Fix: compare projects with and without a coordinator. - **False dilemma, 00:19:17.** "five overlapping products that all work in different ways than like one overly constrained one." Why weak: it omits designs with one lightly constrained product plus sequenced experiments. Fix: state the middle options and why they lose. - **Unsupported quantification, 00:26:22.** "main feed was like 80% of the use." No source, year or definition of "use." Fix: cite the metric and period, or say "a large majority." - **Hasty generalization, 00:22:50.** One computer-use project stands in for how parked projects generally behave. Fix: report how many parked projects were later revived. - **Conflict of interest, throughout.** Both guests speak for a company that sells the models under discussion, and the claim "PMs stay necessary" affects how the company describes hiring. Why weak: it doesn't make the claim false but it shapes emphasis. Fix: pair the panel with independent labor-market or team-performance data. - **Equivocation on "PM," 00:04:22.** The panel shifts between the PM job title and the convenor function. Why weak: the anecdote shows the function was needed, not that a person with the title was. Fix: separate role from title in the argument. #### Tone of voice - Warm, conversational and candid, with audience laughter and a self-deprecating moderator ("I've been cursing too much"). Guests use humor to deflect hard questions, for example the answer to "why didn't you just have Fable do it" leans on a joke about Claude scheduling meetings at 00:05:05 before giving substance. The optimism is consistent, and the concessions (unsolved consolidation, hard trade-offs) sit mostly at 00:17:08 to 00:25:40. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response to accusation or criticism, and this panel has neither an accuser nor a defensive posture. For completeness across the three prongs: **Deny**, no claim is denied. The nearest challenge, "Why didn't you just have Fable do it?" at 00:04:22, gets a direct answer plus a joke and not a deflection. **Attack**, no critics or competitors are disparaged. The mention of OpenAI at 00:24:58 is neutral. **Reverse victim/offender**, no such move appears. Emotional cues are humor and reassurance, not defensiveness. ## 🎯 Conclusion The durable ideas are practices, not forecasts. Name a convenor for connective work. Name a decision owner for each bet. Build shared primitives so agents and the API use the same plumbing. Park what the model can't do, store it as an eval, and re-run it on each new model, as the 2024 computer-use example shows at 00:22:50. The forecasts are less firm. That PMs stay essential rests on one anecdote and comes from guests employed by the vendor of the models under discussion. The 80% Instagram figure and the two-month technology cycle have no sources in this transcript. The panel admits it hasn't solved folding overlapping experiments into one product. **Next steps for readers:** list who plays the convenor and decision-owner roles on your current project, and note what would drop without them. Pick one experiment you shelved for capability reasons and write the eval that would show when to revive it. Get the real speaker attributions, the agent-native write-up author and the model names from primary sources, because the auto-captions garble all of them. ## 🔄 Refactor Appendix — Why Claude can’t be your PM (yet) | Anthropic CPO Panel > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied. The Instagram 80% figure and the model-name banter were dropped as unevidenced. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Claude is not yet a convenor > [!quote] [00:05:05] > I don't think claude is a convenor yet. **Context & use**: - 00:05:05, answering why a human PM was needed when a model was available. Speaker attribution unverified. - The claim is dated and hedged with "yet." Anthropic has a commercial interest in describing model limits and gains, so keep it as a stated position. **Tags**: #xref #articleseed  #### [H2] Building three versions beats debating one > [!quote] [00:07:54] > now it's actually faster to just build three versions and try them out **Context & use**: - 00:07:54, from the explanation of why button-placement craft lost value. - Applies where building is cheap and reversible. Irreversible or regulated decisions still need review. **Tags**: #xref  #### [H3] Shared primitives make agent access native > [!quote] [00:15:41] > a lot of it is really building the right primitives into the product that almost like an infrastructural layer **Context & use**: - 00:15:41, advice for a SaaS company deciding whether to add an agent. - Maps directly to WordPress Abilities API design: one ability layer serving the UI, REST and agents. **Tags**: #xref #share  #### [H4] Overlapping products beat one over-constrained product > [!quote] [00:19:17] > I would much rather have like five overlapping products that all work in different ways than like one overly constrained one **Context & use**: - 00:19:17, the stated preference for exploration before consolidation. - The panel admits at 00:24:58 that folding winners back into one product is unsolved, so this claim carries a known cost. **Tags**: #xref #articleseed  #### [H5] Untested capability conclusions go stale > [!quote] [00:23:33] > you can get capability blind if you try something and then you don't ever try it again or you don't make an eval **Context & use**: - 00:23:33, tied to the parked 2024 computer-use project re-run on each model. - Actionable as a rule: every "the model can't" conclusion carries a date and an eval. **Tags**: #xref #aphorisms  ### 🧠 Atomic Split Pad #### [P1] Store failures as evals **Claim** (one sentence, Courtney's voice): A shelved experiment that leaves behind a runnable eval turns a dead end into a test I can re-run on every new model. **Evidence pointer**: [[#H5]] **Scope & conditions**: Holds when the task can be scored automatically and the environment is reproducible. Breaks when success needs human judgment that an eval can't capture, or when the eval itself drifts from the real task. **Links**: - [[Capability re-testing]] — because extends - [[Bug reports as regression tests]] — because contrasts on where the failure record lives - [[Documentation drift audits]] — because applies-in stale-claim detection  #### [P2] Coordination is the job speed exposes **Claim** (one sentence, Courtney's voice): When building gets cheap, the scarce work becomes connecting people, customers and safeguards, so someone must own that work by name. **Evidence pointer**: [[#H1]] **Scope & conditions**: Supported by one anecdote and stated by a vendor with a stake in the answer. Applies to teams shipping to mixed audiences. Doesn't require a PM title, and small solo projects may not need it. **Links**: - [[Release coordination]] — because applies-in WordPress release cycles - [[Decision owner per bet]] — because extends - [[Vendor claims need independent evidence]] — because contrasts on source of the claim  ### ✏️ Fleeting Queue - [F1] Find the agent-native architecture write-up the panel references (author garbled in captions at 00:12:52) and read it before citing #fleeting #inbox - [F2] Confirm real model names behind "Fable," "Astra," "37" and "Sonnet 45" from Anthropic's model docs #fleeting #inbox - [F3] Open question: who owns provenance when humans and agents both update a shared project view (00:14:59)? #fleeting #inbox - [F4] Article seed: "Say no to the extra tab" on why half of product work is refusing new surfaces (00:26:22) #fleeting #inbox #articleseed - [F5] Test the Abilities API parity idea: list five admin tasks and check each has an agent-reachable ability #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Write a parked-project eval for one shelved experiment and note when to re-run it — DevRel planning, 30 min #things - [ ] Run the human-versus-agent parity check on five WordPress admin tasks and log the gaps — Abilities API review #things  📄 **Full text:** [[Full Document Contents/Articles/Why Claude Can’t Be Your PM (Yet) Anthropic CPO Panel.md|Why Claude Can’t Be Your PM (Yet) | Anthropic CPO Panel by Ami Vora, Mike Krieger, Dan Shipper]] ## Highlights - [00:04:22] happened now is that because we can move faster, that role is like you have to be operationally excellent in that role in a way that like even beyond what you had to be before. But it definitely has not gone away. In fact, I have felt this like renewed like yes, this is a like if you're recruiting for a team and you need people to wear these hats, this hat is increasingly important. >> Can I ask you a serious question though? >> So, >> only serious questions. Serious. This is serious. >> Why didn't you just have Fable do it? [laughter] >> Like, what is And I presumably there's some like Fable 7 you've got access to that we don't have access to yet. So, like what is that person doing that you couldn't just have Fable do? I mean it is interesting like the we have as you can imagine we like run on cloud tag and [00:05:05] we have a lot of like you know cloud provided connective tissue too and it does catch things where you say hey like there's another thing going on in this other part of the organization that you should be aware of you know like we found it via the slash search. So there is that sort of piece as well. I don't think claude is a convenor yet. I had this um I gave a talk a couple weeks ago on like what's the role of the human in the like post super powerful AI age and there's like a bunch of archetypes I was thinking about but one of it was like the convenor like somebody still needs to bring the clouds and the people together to like get the work done and so I think that is the role that I cla organizational pull or the scheduling ability I've yet this is going to be a moment Dan when claude schedules a meeting with me or for me where I'm like who scheduled this meeting it's like oh did I thought you two should talk chat that has not happened yet [laughter] but ([View Highlight](https://read.readwise.io/read/01m3psj97f1j6yt8db1ast4b84)) - Note: [H1] Claude is not yet a convenor - Tags: atom/literature/h1 - [00:07:11] same is true for just models in general. when we were talking Ammy in and in in planning for this conversation, one of the things that you said and you've had a long career in product management um but prior to anthropic is that you spent decades learning how to answer specific kinds of questions as a product leader that you now no longer need to answer. And so I'm curious about okay, what are the skills that you spent such a long time building that you actually don't feel like are that useful anymore? and what are the things that you're leaning on now more and then how does that how does that filter down into the team and your expectations for people that you hire? >> I think it's really true like I spent a really long time trying to understand [00:07:54] how to put myself in the mindset of a user and how something would fit into their lives and like what we should build because it was so expensive to build and ship something. So, you know, literally this would happen where people would have reviews and what we'd talk about is like, am I where should we put this button, you know, and we would talk about the interaction pattern and where like someone's finger would go on their phone and what situation they they would be in where they pulled their phone out of their pocket and what they wanted. And I feel like I tried really hard to get good at that at that mix of like product and design. And I just like I don't think anyone's ever going to ask me that question again. >> That's a fable. That's a fable one. [laughter] >> That's in the past. Like now it's actually faster to just build three versions and try them out and be like, ([View Highlight](https://read.readwise.io/read/01m3psjbx88as0gtp10m3v2v5g)) - Note: [H2] Building three versions beats debating one - Tags: atom/literature/h2 - [00:15:41] okay, do I put an agent in it? Do I make it open and agent native so that cloud can be in it? like how do how do you think about those choices and what a company should care about versus uh leave to leave to a model company like Anthropic to to plug in? Yeah, I think a lot of it is really building the right primitives into the product that almost like an infrastructural layer and you can kind of tell like I I always poke at the whenever I'm using new products um I'll ask it to do something that you know a human could do but you know might be required somebody to think through or architect the thing where everything goes through some shared plumbing that like both agents and like whatever sort of rest API you've built can go through as well and you can tell when they've been thought through that way and it's I have a lot of empathy because if you've been building an application for 20 years just bolting that on is difficult. [00:16:25] So I think is a journey that a lot of companies are going on. But once you get it right, I think that question becomes you can just evolve it over time, right? You can start with, okay, at first I'm just going to have maybe more of the side paint thing. I don't I'm not ready to blow up my whole UI. I'm not ready to create fully generative UIs, but maybe you start experimenting with like, all right, there's also maybe a personal landing page that can be more malleable and is using the same primitives. You're not inventing a whole new sort of infrastructure underneath. So I think that you can evolve your product along the way, but you have like you can tell when things feel super bolted on and like like that's you know feeling like it's not a native part of the product. >> So Ammy I'm curious we're hearing we can go all the way to you know agent native malleable software but uh anthropic you have to serve this whole spectrum of ([View Highlight](https://read.readwise.io/read/01m3psjewvc536yyy1abwbz70q)) - Note: [H3] Shared primitives make agent access native - Tags: atom/literature/h3 - [00:18:34] because, you know, like raised in like the the product like world that I was, I was all about like simplicity and reliability and total predictability because I think that helps people build like a muscle memory and an intuition. But I also think the reality is like we just don't know yet what a lot of things should look like. And so I actually encourage the team to just like take a lot of shots on goal. You know, it's okay to have things that feel somewhat overlapping or duplicative as long as when we figure them out. When we hit product market fit, we go back and we reinccorporate into something that works for everyone. Because I feel like the fail state is you decide from the beginning like based on a theory of exactly what needs to happen and you [00:19:17] like over constrain a product. And I would much rather have like five overlapping products that all work in different ways than like one overly constrained one that we never even gave a chance to actually like meet the market. And how do you and I definitely see this we we do this at every but like I see this all around the industry just doing lots of parallel experiments and seeing which ones one one works and then starting to consolidate as you start to get uh get signal. How do you do that in a way that doesn't feel confusing or incoherent or you know if I'm working on the third team doing this thing like it's can be maybe demotivating like how do you do that? Well >> I think there's a couple of things. one like I think when we were just having this conversation this morning ([View Highlight](https://read.readwise.io/read/01m3psjha4k9gq7nz2y43v7zgp)) - Note: [H4] Overlapping products beat one over-constrained product - Tags: atom/literature/h3 - [00:22:50] change this, you know, thing in some complex software? Like, how do I, you know, do this really complex thing in Photoshop?" And it would get there, but it would get there in like the very like early agent cloud way, which was like this, did that work? Nope, that didn't work. I got to close this. It was like 20 minutes later like, I did it. You're like, I learned nothing. I learn, you know, so like it was neither good for automation nor education. But we parked it and then we would basically just try it with every new model. And we actually, the way we would try it is we actually just had it like running in an eval harness. And the way we actually learned that 37 was like a like a leap in computer use was like just seeing it suddenly took over and then we actually looked at the transcript. We're like wow it's succeeding more often than it's not like this is a real moment. So it is really valuable like we see it as a win in labs if what you did was all if all your project did is show that the models [00:23:33] aren't ready for it yet but you're able to now like externalize that as an eval or sort of like create some uh connectivity to the research team like that's great because it means that 2 3 4 5 6 months from now we'll be able to revisit and hopefully we've learned something along the way um as well but I think you need you need to build early so that you can even have that intuition. >> I definitely agree. I've been calling this like it's it's sort there's a sort of you can get capability blind if you try something and then you don't ever try it again or you don't make an eval and you're like oh the models are just never going to do this and then it just turns out a model comes out like in 3 months and it just completely chang it it improves in ways you never would have thought it would. I'm I'm actually curious if you're here how many of you have used Fable? Raise your hand. Oh, a lot of people are using Okay, this is ([View Highlight](https://read.readwise.io/read/01m3psjkys6ffmqtsg7brpmcya)) - Note: [H5] Untested capability conclusions go stale - Tags: atom/literature/h3