Home › Guides › IETF AI-Agent Drafts
✓ checked 21 October 2026 · source: IETF Datatracker, retrieved 21 October 2026 · by month · by topic · method · download the data
The Internet Engineering Task Force, the body that writes many of the internet’s core standards, approved a working group for AI-agent communication on 8 October 2026, after a surge of proposals. We counted 216 new Internet-Drafts about AI and software agents posted so far in 2026, up from 65 in all of 2025 and 1 in 2024.
216 new agent drafts in 20263.3× the 2025 total39 in the past month1 adopted by a working group
What an Internet-Draft is
An Internet-Draft is a working document submitted to the IETF. Anyone can post one; it carries no endorsement and expires after six months unless it is updated. Most never become standards. A draft gains standing when an IETF working group adopts it, after which the group develops it towards a Request for Comments (RFC), the IETF’s published standards and best practices. Counting new drafts therefore measures how many people are proposing ideas, not how many have been agreed.
New agent drafts by month
Activity rose sharply in the second half of 2025 and again in 2026. 96 new agent drafts were posted in the first six months of 2026 and 120 since the start of July, with the busiest month in July (45). September had 28 and October 26 by the 21st.
| Month | New agent drafts |
|---|---|
| Jan 2025 | 0 |
| Feb 2025 | 0 |
| Mar 2025 | 1 |
| Apr 2025 | 1 |
| May 2025 | 1 |
| Jun 2025 | 4 |
| Jul 2025 | 13 |
| Aug 2025 | 3 |
| Sep 2025 | 1 |
| Oct 2025 | 15 |
| Nov 2025 | 21 |
| Dec 2025 | 5 |
| Jan 2026 | 4 |
| Feb 2026 | 10 |
| Mar 2026 | 29 |
| Apr 2026 | 18 |
| May 2026 | 14 |
| Jun 2026 | 21 |
| Jul 2026 | 45 |
| Aug 2026 | 21 |
| Sep 2026 | 28 |
| Oct 2026 (1–21) | 26 |
Counts each draft once, in the month its first version (-00) was posted. Before 2025, the only agent draft we counted was one from October 2024 on AI-based network management.
The new working group
The IESG, which manages the IETF’s standards process, approved the charter of the Agent Communication Protocols working group (agentproto) on 8 October 2026, according to the group’s Datatracker history. It started as a session at IETF 126 in July. The charter describes an AI agent as “an autonomous, adaptive software system that uses AI models to complete a specific task on behalf of a human user, another AI agent, or an invoking system”.
The group’s scope is narrow. It will define a common “agentic dialog management protocol”: identifiers and lifecycle rules that let a conversation between users, agents and tools be tracked across several hops and services, carried over existing protocols such as HTTP, QUIC or Media over QUIC. The charter says it does not replace application-layer protocols such as the Model Context Protocol (MCP) or Agent2Agent (A2A), which can carry the same context in their own way. It also rules out standardising AI models, agent reasoning, and preventing models from “hallucinating”. Its milestones aim to adopt a use-cases document by December 2026 and send its protocol to the IESG by March 2028.
11 of the drafts we counted already carry “agentproto” in their names, a sign their authors are aiming them at the new group, but none has been adopted by it yet. Across all 282 drafts, only 1 has been adopted by any working group: draft-klrc-aiagent-auth, on AI agent authentication and authorization, in the WIMSE (workload identity) group.
What the drafts are about
We sorted the drafts into rough topics using keywords in their names and titles. The largest group is about identity and permissions: how a website or API can tell which agent it is dealing with, which person it acts for, and what it is allowed to do.
| Topic | What it covers | Drafts |
|---|---|---|
| Identity, authentication and authorization | Proving who an agent is, who it acts for and what it may do (often extending OAuth) | 101 |
| Network operations, routing and 6G | Agents that run or optimise networks, and networks that are aware of agents | 65 |
| Agent communication and other protocols | How agents talk to each other and to tools, including requirements and use cases | 51 |
| Audit, evidence and accountability | Signed records and receipts of what an agent was allowed to do and did | 25 |
| Discovery and naming | Finding agents and naming them, often through DNS | 19 |
| Payments | Agents paying for things, and servers asking them to | 11 |
| Web access, HTTP and crawling | How agents use websites and HTTP APIs, including agents.txt-style policy files | 10 |
Each draft counts once, in the first matching topic. Many drafts touch several topics. Full list in the CSV.
Drafts about agents on the web
For website owners, the most relevant drafts deal with how agents use the web. They include two separate proposals for an agents.txt file, which would let a site declare what automated agents may do, alongside robots.txt; profiles for HTTP APIs built for agents; and a method for rating how “agent-ready” a website is. Related work on identifying automated clients happens elsewhere at the IETF, which we follow on our Web Bot Auth tracker; AI-use preferences in robots.txt are covered in robots.txt AI preferences, and paying agents in x402 adoption.
| First posted | Draft | Title |
|---|---|---|
| 7 Aug 2025 | draft-campbell-agentic-http | A Best Current Practice for Agentic Interactions over HTTP |
| 7 Oct 2025 | draft-srijal-agents-policy | AGENTS.TXT: Strict Policy File for Automated Clients |
| 15 Oct 2025 | draft-cui-ai-agent-discovery-invocation | HTTP-Based AI Agent Discovery and Invocation Protocol |
| 24 Nov 2025 | draft-dhir-http-agent-profile | HTTP Agent Profile (HAP): Authenticated and Monetized Agent Traffic on the Web |
| 9 Dec 2025 | draft-jurkovikj-httpapi-agentic-state | HTTP Profile for Synchronized Resource State (Agentic State Transfer) |
| 8 Jun 2026 | draft-gaikwad-agent-friendly-http-api-profile | Design Considerations and Profile for HTTP APIs Consumed by AI Agents |
| 12 Jun 2026 | draft-car-agents-txt-wellknown | AGENTS.TXT: Capability Declarations for Web Agents |
| 19 Jul 2026 | draft-raskar-agentic-web-federated-resolution | Federated Resolution for AI Agent Discovery Using AI Catalog and NandaIndex |
| 30 Sep 2026 | draft-besleaga-agentic-knowledge-wellknown | The ‘knowledge-linkset’ Well-Known URI for Publishing Knowledge Artefacts |
| 17 Oct 2026 | draft-shah-agent-readiness-evaluation | A Standards-Grounded Method for Evaluating Website Agent-Readiness |
What this doesn’t show
- Drafts are proposals. A draft count says nothing about which ideas will be adopted or deployed.
- Some authors post several closely related or near-identical drafts, which raises the count.
- We matched drafts with “agent” in their name, so agent-related drafts with other names are missed.
- Topics are assigned by keyword and are approximate.
Method
On 21 October 2026 we downloaded every Internet-Draft submission with “agent” in its name and revision -00 from the IETF Datatracker submission API, kept those that were posted, and counted each draft once by the date of its first version. That gave 282 drafts from 2024 to 2026. We read the abstracts of all drafts whose title did not clearly mention AI or autonomous agents and removed one that is not about AI or software agents (the SSH key-agent protocol). Older “agent” drafts, for example on DHCP relay agents and Diameter agents, date from 2007 to 2016 and are excluded. Working-group status and abstracts come from Datatracker document records; the charter and its history come from the agentproto group page. We use only titles and metadata; draft texts are copyright the IETF Trust and their authors.
Download the data
- IETF AI-agent Internet-Drafts (CSV): 282 drafts with title, date first posted, rough topic, working-group adoption and Datatracker link.
Cite as: ProxyPicker, “The IETF’s AI-Agent Draft Surge”, 21 October 2026, https://proxypicker.com/ietf-ai-agent-drafts/. Data: IETF Datatracker.
Related: Web Bot Auth tracker · AI-crawler census · MCP server SSRF.
Press & data requests: email [email protected] for interviews, the raw data behind this page, or corrections. See also our contact page.