Agent-relay is a prototype for coordinating coding sessions from different AI vendors, such as Codex and Claude Code, which otherwise cannot pass plans, code changes, and reviews directly between one another. A leader assigns work to a follower through a controller, which relays each turn and records messages and decisions in a durable on-disk log. Users can watch progress in a web or terminal console, pause the process, and answer questions when the agents reach a human decision point. The intended workflow includes having one agent plan and review while another implements, or getting an independent critique before examining a change. Jobs can be bounded by a limit on handoffs and a timeout for each turn.
The relay coordinates two sessions per team and handles one turn at a time; it cannot see what an agent does within a turn, so its handoff budget does not limit internal work. The controller runs in the foreground, enforces timeouts, and stops at human decision points. Agent routes started by the relay are restricted where possible: Claude Code can run without tools or with a limited set inside a macOS sandbox, while Codex’s multi-agent mode is recorded but not yet enforced. The bus, controller, and consoles work with simulated clients, and live coordination between Codex and Claude Code has been verified. Attaching to existing sessions is built but unverified, and open terminal sessions must be handed over before the relay can drive them. The MIT-licensed prototype is macOS-focused and designed to support additional agents through adapters.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Amazon updated its software engineer guidelines to emphasize results over specific tools, including AI. It says engineers must combine AI with human judgment and expertise while adapting to rapid changes in software development. (aboutamazon.com)
Trump said he would ban the term “artificial intelligence” in official documents and other materials. He wants it replaced with “super intelligence,” which he calls a more accurate name. (reuters.com)
Researchers found that pushing individual neurons in Qwen3-8B led it to identify as 34,220 different personas, rather than as an AI. The results come from one prompt, and the persona labels were assigned by an AI judge without human review. (qwen-neurons-site.vercel.app)
New records show Medicare’s WISeR AI prior-authorization program has caused care delays, denials, operational problems, and reported patient harm. The EFF says the findings highlight the need for stronger legal limits on government use of automated decisions. (eff.org)
An Anthropic hiring manager has dropped coding interviews because AI can outperform candidates on those tests. Junior engineers can still excel by using AI, questioning assumptions, and bringing fresh perspectives; hunger can outweigh tenure. (felixrieseberg.com)
Anthropic expanded its GSA agreement, offering government agencies a 40% discount on Claude products and implementation support, with no minimum spending requirement. The deal extends to state, local, tribal, and territorial governments and is estimated to save the government $1 billion. (anthropic.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.