Right now, there are 11 different IETF drafts trying to answer the same question: how should AI agents discover and interact with websites?
The front-runner, agents.txt, expires on April 10, 2026. Two days from now. And none of these proposals talk to each other.
I spent the last week reading all 11 drafts. Here is what I found.
What is agents.txt?
Think robots.txt, but for AI agents. Submitted by Srijal Poojari as draft-srijal-agents-policy-00, it proposes a file at /.well-known/agents.txt where websites declare policies about agent interaction: which agents are allowed, what they can do, rate limits, authentication requirements.
It is the simplest proposal on the table. That is probably why it got traction first.
All 11 drafts, compared
Here are all 11 competing drafts, sorted by expiry date:
| Draft | Backer | Mechanism | Expiry |
|---|---|---|---|
| agents.txt | Srijal Poojari | File-based (/.well-known/) | 2026-04-10 |
| ACDP | Command Zero | DNS-based capability records | 2026-06-30 |
| agents.md | Community | Markdown-based policy | 2026-07-01 |
| AID | Community | Decentralized identifiers | 2026-07-15 |
| ADS | Community | Service discovery | 2026-07-30 |
| ARDP | Cisco/AGNTCY | Resource discovery protocol | 2026-08-05 |
| AWP | Community | .well-known/agent.json | 2026-08-10 |
| BANDAID | Community | Bridge protocol | 2026-08-20 |
| ANS | Solo.io | DNS + PKI | 2026-09-01 |
| Agent Networks | AAIF | Network mesh | 2026-09-15 |
| MCP Server Cards | MCP Community | Capability metadata | N/A |
Eleven different answers to the same problem. Five use .well-known endpoints. Three use DNS. Two propose entirely new protocols. One is just markdown in a file.
Why the April 10 expiry matters
When agents.txt expires, the IETF removes it from the active draft list. Anyone who built against it loses their reference spec. The draft can be resubmitted, but that takes weeks, and there is no guarantee the author will bother.
Meanwhile, every AI company is doing their own thing. OpenAI has one pattern. Anthropic uses another. Google does something else again. Without a standard, website operators cannot set reliable policies for agent interaction.
Here is the part that makes this uncomfortable. The EU AI Act expects compliance with website policies. Upcoming US regulations point the same direction. But compliance with what format? Nobody knows, because nobody agreed on one.
Three scenarios after expiry
Renewal. The author resubmits a revised draft. Probable, but slow. Fresh review cycles, new comment periods.
AAIF takes over. The Artificial Agents Interoperability Foundation (146 member companies) produces a unified proposal. Likely long-term, but we are looking at 12-18 months minimum.
Fragmentation. The industry splinters into 3-4 incompatible walled gardens. Companies adopt whichever draft their largest vendor ships.
Honestly? We are already in scenario 3. These are not just proposals sitting in a queue. Cisco is building ARDP into their agent infrastructure. Solo.io is shipping ANS. The MCP community has their own approach and it is gaining adoption independently.
What I would do if I were building an agent today
Do not pick a winner. Implement agents.txt because it is the simplest and most referenced, but also:
- Parse
/.well-known/agent.jsonif it exists (covers AWP) - Check DNS TXT records for capability declarations (covers ACDP and ANS)
- Support MCP Server Cards for capability metadata
Build a translation layer between formats. When a winner emerges (or when AAIF ships their unified spec), you swap the backend. Your agent keeps working either way.
The full comparison table, live countdown timer, and links to all 11 drafts are on the analysis page. I will keep updating it as drafts expire or get renewed.
Top comments (0)