Loading today’s stories
Product1 distinct publisher2 min readPublished
PostHog cannibalized its agentic Code product: self-driving shipped inside Web, the leftover shell became Desktop, and the company says Desktop still has no product-market fit.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
The arithmetic on the disappointment score deserves doing before anyone borrows this playbook. PostHog put the Sean Ellis question to its own team before a recent launch and got 23% [10], against the roughly 40% benchmark it cites [11]. That is 17 points short, and about 58% of the bar [16]. Publishing it is more useful than the six rules wrapped around it, because it puts a floor under what a pre-fit internal number looks like at a company that sells measurement tooling.
The structural advice is narrower than it first sounds. Three to five people, reporting to a founder rather than to the department they are disrupting [5]. The reporting line is the mechanism: the department being cannibalized owns both the revenue and the argument for stopping you, and PostHog also tells teams to expect lower margins, worse retention and more pivots than planned [7]. Those numbers, reported into the org that owns the flagship's margin, read as failure. Reported to a founder, they read as a seed round.
PostHog's own comparison case is Vercel, which ran its agents in production and shipped the gaps as products: AI Gateway, Chat SDK, the Agent Stack [14]. That is a different move. Shipping the gap adds surface; de-bundling subtracts it, and subtraction is the cheaper experiment, because you find out which half of the product users were actually holding onto. PostHog ran Code as a closed beta with real users, and what came back pushed its core ideas into Web [15] rather than into a bigger Code. The company also says it is fine with customers who never open the web app at all, arriving instead through the MCP, the Slack app, Desktop or the CLI [13].
The threat model behind this is stated plainly: frontier labs moving onto your territory, plus small teams shipping "[Your App] for 2026" [17]. The de-bundling result is the closest thing in the account to evidence. One feature survived contact with an installed base and one container did not, which is a concrete answer to what an AI-native team would skip if it rebuilt your flagship from scratch.
Ranked by verification strength, evidence, and original report placement.
PostHog's in-house disruption effort started as PostHog Code, an agentic coding tool and product editor.
The most valuable thing that grew inside PostHog Code, self-driving, was de-bundled and shipped into PostHog Web.
The container Code left behind became PostHog Desktop, built around multiplayer spaces, generative UI artifacts, and the thesis that shared business context is key to a product that builds itself.
PostHog Desktop is in open beta and, according to the author, does not have product-market fit yet.
PostHog ran the disappointment question before a recent product launch and scored 23%.
PostHog cites an industry product-market-fit benchmark of roughly 40% on the disappointment question.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single-source first-party, specific but unverifiable
Everything rests on one vendor-authored newsletter post. Its strength is specificity and self-incrimination: named product transitions, a disclosed 23% internal survey score, and an explicit admission that Desktop lacks PMF. Its weakness is that nothing is independently corroborated — no second publisher, no named customer, no methodology for the survey, and no citation for the ~40% benchmark or the Vercel precedent.
Pre-PMF: one shipped flagship feature, one open beta
Real shipping has occurred — self-driving landed inside PostHog Web and Desktop reached open beta — plus one disclosed customer building on the MCP with Claude and internal instrumentation watching Desktop sessions. But no user counts, retention, revenue or beta-size figures are given, the company states Desktop has no product-market fit, and its own internal disappointment score sits well below the benchmark it cites.
Understated for a vendor post
Negative because the claims are more restrained than the underlying facts would permit. A marketing-owned newsletter leads with 'doesn't have product-market fit (yet)', volunteers a 23% disappointment score against the ~40% benchmark it names, calls the result 'painful', and describes its own product as work-in-progress. The one place framing runs ahead of evidence is calling a single unnamed customer prototype proof that 'the demand is proven', which caps how far negative this can go.
Vendor-authored product marketing
The only source is PostHog's own newsletter, written by the person whose stated job is marketing the company's in-house disruption effort. The post promotes named PostHog surfaces (Desktop, Web, MCP, Slack app, CLI) and recruits readers into an open beta, and it cites third parties (Vercel, GitHub Copilot) to validate its own strategy. Mitigating factor: the disclosure of an unflattering internal score and a no-PMF admission runs against the promotional incentive.
Moderate — reliable on self-report, blind beyond it
High confidence in what the company says about its own product decisions and internal survey result, since a vendor is authoritative on its own shipping history and has little reason to invent an unflattering number. Low confidence on anything outside that: the customer prototype, the ~40% benchmark, the Vercel description, and any statement about how Desktop is actually performing in open beta. Single-publisher clusters cap this dimension.
product
A 2x LLM bill is not a bug report: token spend is an observability problem1 distinct publisher
build
Anthropic's Browser Use hands Claude element refs, and hands you the browser1 distinct publisher
product
Thomson Reuters spent $40M to make a $450K training run worth doing2 distinct publishers
build
Amazon Q executed code from any repo you opened, and it is not the only one1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 25, 2026