Skip to main content
usedby

How Harvey runs marketing with agents, and what its stack shows

Harvey makes product films with agents in Figma, ElevenLabs voices and a Claude artifact as tracker. What its own account shows, and what to copy.

On 21 October 2026, Harvey published how its marketing team now works with agents. The post is signed by Shawn Farsai and co-founder Gabe Pereyra. It opens with a question Figma asked Harvey: what was it building with Figma's MCP server? The post is "the long answer". (Harvey blog)

Harvey's CMO, Rachel Hepworth, gave the short version on LinkedIn the same day. Figma "reached out recently because our usage of their MCP server had skyrocketed." She adds that the team had tried tailored demo films six months earlier and "couldn't get it to scale". New capabilities in Figma and newer AI models changed that. (Rachel Hepworth on LinkedIn)

This is a first-party account. It names each tool, says what the tool does in the pipeline, and shows where the work broke. A vendor logo wall carries none of that.

What Harvey built

The first project is a launch film for Harvey Artifacts. Harvey says the team made it "entirely with agents, inside of Figma". Each scene is a Figma Motion frame, and the cuts export straight from the file.

The second project is about scale: a tailored product film for each of the hundreds of accounts Harvey wants to win. The first prompts came back "half decent". Each film then needed 15 to 20 minutes of back-and-forth, and the team saw that this would not scale.

So Harvey built a foundation first. Agents crawled the app and rebuilt every key screen as Figma components, a library called Product Map. A second library, Motion Primitives, holds reusable animated parts. On top sits a menu of 44 company-agnostic demo flows, each built around one moment buyers ask to see.

The pipeline, step by step

For each account, agents use Enterpret's MCP to pull together Gong calls, Salesforce win-loss reports and Slack. An agent then writes the script, with a fictional company on screen.

The voice comes from ElevenLabs. Harvey built its narrator's voice from ten minutes of her footage, and agents call ElevenLabs through its MCP.

An engine of about 750 lines of JavaScript turns each script into a film inside Figma. It handles the camera, transitions, captions, the pointer's click rhythm and the end card. It then lays out contact sheets and flags clipped text, small type and stray pointers, before anyone watches.

When a frame needs a photo, a texture or a shot, the agent uses Figma Weave, which is now part of Figma. It reaches Weave through the same Figma MCP it designs with.

Everything runs through one tracker: "a Claude artifact that every agent and person reads and writes". Each film has a row with its account, story, status and links, and agents update their own rows.

The stack, and what usedby could see before

Tool
ToolRole in the pipelineProof on usedby
FigmaRole in the pipeline
Brand library, product screens, motion (Figma Motion), film assembly, generated…
Proof on usedby
Confirmed, new line
ElevenLabsRole in the pipeline
Every line of voiceover; its voice library holds clones of the narrator and teammates
Proof on usedby
Confirmed, was observed
EnterpretRole in the pipeline
MCP that pulls account context together for each script
Proof on usedby
Confirmed, new line
GongRole in the pipeline
Sales calls used to tailor each account's film
Proof on usedby
Confirmed, new line
ClaudeRole in the pipeline
The shared tracker, a Claude artifact read and written by agents and people
Proof on usedby
Confirmed, second source
SalesforceRole in the pipeline
Win-loss reports pulled for each account
Proof on usedby
No usedby page yet
SlackRole in the pipeline
Account context, and the channel where Harvey's internal agent Spectre takes…
Proof on usedby
No usedby page yet

Before this post, usedby showed ElevenLabs at Harvey as observed: the vendor listed Harvey on its own pages. Figma, Gong and Enterpret had no line. Claude was already confirmed by a customer story on Anthropic's site.

A company describing its own use is the strongest proof usedby records. The post moves ElevenLabs to confirmed and adds Figma, Gong and Enterpret at that level. On 22 October 2026, the Harvey page on usedby lists 22 tools: 8 confirmed, 8 observed and 6 signals.

A logo says Harvey is a customer. The post says every line of voiceover in Harvey's films comes from ElevenLabs, and why the team chose Figma over HTML for layout. That is the difference between a listing and an account.

MCP is the integration layer

The term that comes back most in the post is MCP. Figma, ElevenLabs and Enterpret are each reached the same way: an agent calls the tool's MCP server.

Harvey ties its output to that layer. In the 14 days before the post, Shawn Farsai used 15.6B tokens across Figma, ElevenLabs and Weave. The closing line reads: "MCP is what makes that possible."

Figma used like a codebase

"We think about our Figma the way engineers think about a codebase," the post says. Harvey's Agentic Brand Guidelines act as the top-level AGENTS.md. Each project folder has its own AGENTS.md page that adds only what that project needs.

The team started with markdown files. They worked for design tokens and broke one step up, for example on how to build a blog chart or a long PDF guide. The guidelines became a Figma library with worked examples, so agents see the relationships between elements instead of reading about them.

Layout settled the choice of canvas. Harvey tried HTML, including a layout engine built on Pretext, but a headline's cap height never sat exactly on a grid line. Figma trims text to cap height and baseline, and holds each grid as a layout agents apply.

Where it broke

Only one agent can work in a Figma file at a time. Agents queuing in one shared file pushed calls to two minutes, and the calls timed out.

The fix was a bus: three lane files, each with its own builder agent, plus a coordinator for export, mixing and delivery. One batch of 36 films took 69 minutes.

Then the team hit its daily Figma call cap. Overnight, agents scripted and voiced the next 97 films.

What a marketing leader can copy

Start from a storyboard. Harvey treats it like a PRD for film: every shot, line and transition is decided before production starts.

Watch the back-and-forth. Fifteen to twenty minutes of edits per asset was Harvey's signal to build shared libraries before making more.

Give agents brand rules they can see. Colors and type fit in a file. Layout and composition fit better in a library with worked examples.

Connect account data at the source. Harvey's scripts draw on sales calls, win-loss reports and Slack, pulled through one MCP.

Check the output by machine first. Contact sheets that flag clipped text and stray pointers catch the obvious errors before a person reviews.

Keep one tracker for agents and people. One row per asset, updated by whoever, or whatever, moves it forward.

What the post does not say

The post gives no cost, no conversion figure and no result for the tailored films. It does not say how many films reached accounts. usedby records which tools Harvey uses and how; it did not measure any outcome.

Source: How Harvey is Building a Frontier Marketing Org, Harvey, 21 October 2026. Short version: Rachel Hepworth on LinkedIn, 21 October 2026. Tool lines: Harvey on usedby.

Sources

  1. Harvey blogharvey.ai
  2. Rachel Hepworth on LinkedInlinkedin.com

Questions

Which tools does Harvey's marketing team use for its agent-made films?

According to Harvey's post of 21 October 2026: Figma (with Figma Motion and Figma Weave), ElevenLabs for every line of voiceover, Enterpret's MCP to pull Gong calls, Salesforce win-loss reports and Slack, and a Claude artifact as the shared tracker.

How fast did Harvey produce the tailored films?

One batch of 36 films took 69 minutes. When the team hit its daily Figma call cap, overnight agents scripted and voiced the next 97. Harvey gives no cost or conversion figure.

What proof level does usedby give these tools at Harvey?

Confirmed, because Harvey describes its own use. Before the post, ElevenLabs was observed (listed by the vendor), and Figma, Gong and Enterpret had no line.

Tools in this article