Agentic advertising

The term has two uses. The first, and the one this entry focuses on, is an AI agent doing the media buyer’s work: reading a balance, drafting a campaign, requesting a price, submitting it and pulling reports. The second is advertising that appears inside agents. Gravity, for example, describes itself as reaching people in conversations with LLMs, agents, chatbots and AI search, and OpenAI is testing Sponsored Agents, an ad format in which people chat with an AI representative of a business.

For agent-run buying, the platform has to be readable by machines. Admux publishes agents.txt with task instructions, an OpenAPI description of its HTTP API, and an MCP server that mirrors the same operations. An agent works through the same steps a person would: list channels, upload creative, create and validate a draft, request an offer, submit, then launch, pause or report.

The safety model is what separates agentic buying from handing over a password. In Admux each agent gets its own key with scopes and a spending limit, mutations that move money require an Idempotency-Key, and payments, refunds, membership and key administration stay human-only.

Start an agent with read scopes and reporting, then add drafting, then reserving funds, and grant launch last. In Admux the default launch mode is review: approved legs wait for a person to launch them, and an agent can launch only with a key scoped to allow it. Remember that a submitted campaign is not a live campaign; Admux separates acceptance, approval and serving, so an agent should poll the operation and campaign status rather than assume success. Pause and cancel are requests, not confirmations. Gravity’s own dashboard MCP server, by contrast, runs actions without confirmation and has no spend limit, which is worth knowing before connecting an agent to it directly.