Half of Linear's work is agent-created, and Salesforce says its UI isn't the product

Two disclosures this week point at the same shift: Linear says agents now create about half the work inside its product, up from 3% a year ago, and Salesforce has started shipping a version of itself that you never open. Between them they change what a software company is actually selling, and therefore what its marketing team is supposed to say.

The Linear number appeared in the blog post accompanying a $99 million secondary at a $2.5 billion valuation, and was pulled out by SaaStr's Jason Lemkin, who writes as a long-time investor in this category. Alongside it: agents are installed in 95% of paid Linear workspaces, and issues with a pull request attached have grown sevenfold since the start of 2026. Lemkin's caveats travel with the figure. Linear reported the share of work agents create, not half of all issues, and an install count is not an engagement count.

The Salesforce half is Claudeforce, its expanded partnership with Anthropic. The first product, Salesforce in Claude, lets people query CRM records, update them, review deals and run workflows without touching the Salesforce interface. MarTech reports it launched with 37 prebuilt sales skills to pilot customers, an open beta planned for September and marketing capabilities to follow. The groundwork was Headless 360 in March, which exposed the same data to agents over APIs and MCP servers.

The interface was never the moat, but it was the marketing

Salesforce president of applications and marketing Patrick Stokes put the repositioning plainly to VentureBeat: "The value of Salesforce is not in our UI itself. It's not the application." The value, he said, sits in the data and the metadata. Marc Benioff gave CNBC the defensive version, calling the "SaaSpocalypse" narrative nonsense because frontier models depend on CRM.

Both statements are probably true, and both are inconvenient for marketers. For twenty years the demo was the pitch: screenshots on the pricing page, a guided tour in the trial, an onboarding flow built to produce a first-session aha. If a meaningful share of the work inside your product now arrives through an API, an MCP server or someone else's chat window, that apparatus is aimed at a shrinking fraction of your usage. What gets evaluated moves to your data model, your permission system and your workflow logic — and almost nobody has marketing pages for those.

The demo is no longer the product tour. It is the surface an agent sees.

Seats held. Storage did not.

The fear was that agents would drain the interface and take seat revenue with it; the reported numbers do not show that yet. Atlassian said 98% of the people using its MCP server were also active in the Jira UI in the same month, and that MCP adopters expand paid seats faster and grow ARR at twice the rate of non-adopters. Linear held 177% net revenue retention above $100 million in ARR through the exact period agents went from 3% to 50% of created work.

What moved is consumption. SaaStr, which runs its own operation with three people and more than twenty production agents, describes going from roughly 5GB to roughly 40GB of Salesforce data in about thirty days — around 21 million records, 99% written through the API — and being flagged for storage overages during a week nobody had logged in. That is the pricing story hiding inside the adoption story. Agents consume more of your product, faster, on dimensions your packaging may not meter.

The metric asymmetry is about to become a marketing problem

Lemkin makes a point worth stealing for your competitive work: Linear published a composition number, Atlassian published a growth rate. Jira work items and Confluence pages created through its MCP server were up nearly fourfold quarter over quarter, and total MCP calls up more than 400% — impressive shapes off an undisclosed base. monday.com and Asana published neither, offering cumulative activity totals with no denominator.

Growth rates are what you publish before you have arrived; composition is what you publish after. Once one credible vendor in your category says what fraction of its usage is agent-generated, every competitor's percentage-increase claim starts to read as an evasion.

What a software marketer should do differently

• Instrument the composition metric now, before a competitor sets the disclosure standard for your category.
• Audit pricing against agent behaviour. Storage, records, API calls and per-agent identities are where the bill moves; seat contraction is not supported by the current data.
• Give your data model, permission inheritance and workflow logic real marketing pages. Those are the evaluation criteria now, and most vendors still bury them in developer docs.
• Treat integration partnerships as higher risk. ServiceTitan cut off Podium after nine years and roughly a thousand shared customers, on thirty days' notice, once Podium's agents began holding the customer record.

One caution against overreacting. Salesforce demonstrated Claude assembling a custom dashboard from CRM data on the fly, then said explicitly that the demonstration is not a product feature. Agent usage is real and measurable; agent-native buying is not yet. Build for the first without pricing for the second.