Right now, if an AI agent needs to book you an appointment or request a refund on your behalf, it has to scrape a website, fill out a contact form designed for humans, or hope an API exists. There's no standard protocol for agent-to-website communication the way robots.txt handles crawler permissions.
Agent.json is an open protocol trying to fix that. It works through three endpoints that any website can implement:
GET /.well-known/agent.json- Tells agents what actions are available, what data format to use, and how to authenticatePOST /.agent/inbox- Receives structured messages from agentsGET /.agent/inbox/:id- Lets agents check back for responses
The protocol uses JSON Schema to define exactly what data agents should submit for each action, so there's no guessing. It supports synchronous responses, polling, and HMAC-signed webhook callbacks for longer operations. A reference implementation already runs on Cloudflare Workers, and the spec is platform-agnostic - it can be added to Express, Django, Rails, or any other framework.
This sits in interesting territory. AI agents are getting good enough to handle real tasks, but the web isn't built for them. Every website has its own contact forms, its own checkout flows, its own way of handling requests. Agent.json is betting that businesses will want to make themselves "agent-accessible" the same way they made themselves search-engine-accessible with sitemaps and structured data two decades ago.
The obvious challenge is adoption. Robots.txt worked because search engines had enough power to make webmasters care. AI agents don't have that leverage yet. But as tools like ChatGPT, Claude, and Google Gemini add more agentic features, a standard like this could save everyone from building one-off integrations for every website.