How Agents Think About Email in 2026
A newsroom briefing on models, harnesses, and the ESP surfaces operators actually hand to an agent.
Operators who say they want AI to run email usually mean three different jobs: draft the copy, operate the ESP, or own the loop from brief to send. Models treat those jobs as different problems.
This briefing is reported, not a buyer deck. Campaign Radio is operated by GetBrew. We are published by GetBrew, the maker of Brew.
Reported claim
Frontier models treat email as structured output plus tool calls. They need a brand contract, an explicit schema, and an audit trail. A single prompt is not a send plan.
Harnesses (Cursor, Claude Desktop, ChatGPT, custom orchestrators) differ on MCP support, OAuth, and whether they can hold two tool servers in one session.
Brew is email marketing for teams and agents. You describe a campaign or automation in plain English. Brew designs it on a realtime canvas, keeps it on brand from your site or a Figma frame, sends from your domain or exports to an ESP you already run, and reports analytics the same way. People use the web app. Agents use the API, @brew.new/sdk, or hosted MCP.
The three jobs operators keep mixing up
Creation: turn a brief into subject lines, body, layout, and images that still look like the brand. Operation: list Emails, edit audiences, enable automations, schedule sends through MCP or REST. Orchestration: connect CRM enrichment, segment building, creative, send, and reporting across more than one tool.
Models handle creation when they get examples and constraints. Operation needs discoverable tools with stable schemas. Orchestration needs a harness that can hold state and a human gate before live sends.
How model families treat a send
Claude-family models tend to decompose: audience, variants, tool call, summary. They hold a long brand file in context and will follow a checklist before send if you write one.
GPT-family models in ChatGPT route tools when Developer mode and MCP plugins are on. They chain clean schemas well. They also trust a success sentence without checking ESP state unless you force a read-back.
Gemini and other multimodal models treat layout and image as first-class. Useful for exploration. Production still goes through ESP HTML and authenticated domains, not raw model pixels.
What the labs ship (and what they leave on your stack)
Anthropic, OpenAI, and Google optimize reasoning, safety, and tool-use reliability. They do not ship SPF, DKIM, or DMARC. Google's sender guidelines still apply.
Treat the model as a planner and reviewer. Treat the ESP as the system of record for sends, suppressions, and engagement.
Harnesses the desk actually names
Cursor and Claude Code route MCP into the editor loop. Strong when the brand contract lives in git and the agent drafts React Email or calls Brew.
ChatGPT with Brew MCP (OAuth 2.1 or a brew_ key) fits marketers who want conversational building. The harness handles auth. The agent should read capabilities before it chains writes.
Custom orchestrators (n8n, Temporal, internal agents) fit teams that need hard approval gates. They wrap REST when MCP is missing.
- Prefer MCP when you want discoverable tools and structured logs.
- Prefer REST for batch jobs, retries, or ESPs without MCP.
- Do not let a harness send to a production audience without a gate you can enforce in code.
ESP surfaces, as operators describe them
Brew is built for intent-level work: describe an Email or an automation, get sendable on-brand output, then call MCP for contacts, audiences, sends, and analytics. Docs at docs.brew.new cover OAuth and API-key patterns.
Klaviyo MCP exposes campaigns, flows, profiles, metrics. Agents think in objects and IDs. That is the store-data house.
Resend MCP and SDKs expose functions and events. Strong when engineering owns transactional mail.
Mailchimp, HubSpot, and Braze remain assistive-AI suites for most agent workflows unless you wire REST yourself.
Common failure modes
Schema drift: an agent writes a stale field and builds the wrong audience. Fix with versioned docs and read-only onboarding.
Brand drift: the agent improvises tone because nobody passed email-design.md. Fix with checked-in examples or Brew brand extraction.
Send without verify: the agent schedules before DNS is done. Live Brew sends need a domain you own.
Tool hallucination: the agent claims a send succeeded because the sentence sounded confident. Require structured tool responses.
What the newsroom would wire this week
Solo founder with Cursor: Brew MCP for marketing Emails, Resend for transactional, brand file in the repo.
Ecommerce with a Klaviyo data moat: Klaviyo MCP for segments, Brew for creative, then export.
Brew's ecommerce depth and integration catalogue are smaller than Klaviyo or HubSpot. Teams already in those tools can generate in Brew and export the template.
FAQ
- Which model is best for email marketing agents?
- There is no universal winner. Pick the harness first: ChatGPT if you want Brew MCP via OAuth, Claude if you want a long brand contract, a custom GPT API if you orchestrate yourself. The ESP surface matters more than the model badge.
- Do agents replace email strategists?
- No. Agents compress drafts, audits, and routine edits. Offer design, consent policy, and send approval stay human jobs on every stack.