Archive

Proton's chief technology officer has published an engineering memo saying developers should direct AI models and rebuild the whole development process around them. It is a statement of intent: no numbers, no timeline, no named tools. Watch it as a checklist of questions, not as proof.

Claude Code 2.1.314 (30 October) adds a setting that makes its file-editing tools refuse any path outside the working folders, whatever permission mode is on. It does not limit shell commands. Adopt it if you run with relaxed permissions, and add a sandbox for the rest.

Claude Code's changelog says fast mode for the Opus 4.8 model is deprecated as of 30 October and will be removed on 30 November. After that, requests fail instead of falling back to the normal model. Flag it: check shared settings and automation, and move to Opus 5.5.

On 22 October 2026 the IETF's steering group approved the charter of DAWN (Discovery of Agents With Names). It will standardise only the first lookup of an AI agent's minimum public information, and excludes capability catalogues, internet-wide search and agent identity. One milestone is dated June 2027.

The project behind the Personal Agent Protocol published draft 0.2 on 24 October 2026 and says each 0.x version can break the one before. Draft 0.3 is already in progress with its own breaking changes. Do not build against it yet.

Google's open-source Agent Development Kit (ADK) version 2.13.0 added a classifier that routes agent workflows using TypeSafe's Jev model. Jev picks answers from fixed options instead of writing text, and is in early access. Its speed and price claims are the company's own.

Version 1.8.0 of AGNTCY's Agent Directory, an open-source registry of records about AI agents, came out on 15 October 2026. Publishers can sign identity claims that are re-checked against current keys, and operators can run content rules in a report-only shadow mode before enforcing them.

On 12 October 2026 the maintainers proposed Agent Plugins to the Agentic AI Foundation as a Growth-stage project. It bundles reusable agent components into one installable package. The foundation's technical chief questioned its licence, and the proposal has since passed its technical committee.

On 14 October 2026 a maintainer recorded that MCP's core maintainers had accepted SEP-3398, which moves the protocol to two specification releases a year. It merged on 27 October. The next revision is proposed for March 2027.

agentproto is an IETF working group, approved in October 2026, on how AI agents keep track of one dialog as it passes between systems. It does not replace existing agent protocols or define a new transport. It has a charter, not yet a protocol.

The Model Context Protocol project marked SEP-2127 as Final. It defines a Server Card, a small public description of a remote MCP server that a client can read before connecting. It is an extension; the core specification did not change.

On 6 October 2026 Sierra announced the Personal Agent Protocol, an open standard it is developing with Meta for how a person's own AI agent interacts with businesses. It is built on OAuth, the standard for delegated sign-in. The post gives no adoption figures.

A. N. Besleaga proposes that a site publish one list of its knowledge files at a fixed web address, each link able to carry a fingerprint so a client can check what it downloaded. A first draft by an independent author, intended as Informational. Not a standard.

Version 1.2.0 of the official Python library for A2A (Agent2Agent, a protocol for AI agents talking to each other) came out on 29 September 2026. It adds caching headers to the Agent Card endpoint and serves REST responses as application/a2a+json. Plumbing, not a new capability.

A proposed IETF working group, DAWN (Discovery of Agents With Names), would standardise only the first step of finding an AI agent: asking for its minimum public information. On 30 September 2026 its charter went to a first ballot. A proposal then, since approved on 22 October.

Agent Registry Protocol (ARPA) is one author's IETF proposal for how a registry could say who runs an AI agent and what it may do, kept strictly apart from whether the agent exists. It is an individual draft, not a working group document. Useful vocabulary, too early to build on.

C. Drake of 1id.com published revision 04 of an IETF proposal that gives each AI agent a permanent, global identifier, with five trust levels from hardware-backed to software-only. It is an individual Experimental draft. It asks for one central registry and cannot revoke an identity.