What 1.46 Million Agent Tool Calls Reveal About Content Operations

Sanity's June 15 analysis of agent activity offers a useful alternative to predictions about the future of CMS work: examine recorded operations. The company reports 1.46 million MCP tool calls between September 2025 and April 2026, involving 12,300 users across 12,500 organizations.

Within that window, monthly calls grew from 7,400 to 521,000. Sanity says 91% of activity covered daily work such as querying, editing and publishing, while migration represented 3% and localization 2%. These are figures from Sanity's observed users and tooling, not a measurement of the whole CMS industry.

A tool call records that software requested a unit of work. A complete editorial task can require many calls, and a small number of calls can sometimes accomplish a substantial change. Counts are useful, but they do not directly measure completed projects, business value or time saved.

That distinction matters when teams compare their own experience with a vendor's dataset. A workflow that makes fewer calls may be more efficient, or it may simply cover a smaller task. A high-volume user may be productive, or may be repeatedly querying material that should have been fetched once.

Our analysis is that organizations should connect operation records with accepted outcomes. Which drafts were approved? Which migrations completed correctly? Which queries helped an editor make a decision? Those connections turn activity into evidence a team can use to improve its process.

Sanity describes an optional intent field added to its tools in March 2026. The resulting explanations can help classify why a request was made. The company also interviewed twelve enterprise content leaders across eight industries to add context to the quantitative data.

Intent metadata is valuable, but it is an agent's stated purpose rather than independent proof of the result. A request can describe a useful goal and still produce an incomplete change. Interviews add depth, while a small qualitative sample should not be treated as a statistical picture of all content teams.

For a local pilot, this suggests recording the task, the affected material, the tool results and the review outcome together. The record need not be elaborate. It should be sufficient for a person to understand what was attempted and whether it was accepted.

The prevalence of daily operations is interesting because it shifts attention away from spectacular demonstrations. Finding content, correcting a field and preparing a publication are ordinary tasks with clear outcomes. They can be easier to evaluate than an open-ended request to redesign an entire site.

An organization might start with one recurring audit, measure accuracy and review effort, and decide whether to expand. A rare migration can still justify substantial investment, but its value needs a project-specific explanation rather than a share of tool-call volume.

The dataset supports a measured conclusion: agent interaction with a CMS is already observable production activity for the users studied. Understanding its value requires the next layer of evidence: the quality of the work, the human decisions around it and the outcome that remained useful after the calls ended.