A bunch of large banks from around the world have published a list of principles for trusted agentic commerce. It’s short, sensible and reasonably readable. Their first principle poses the crunchiest question:
“1.1 ๐๐ญ๐ญ ๐ฑ๐ข๐ณ๐ต๐ช๐ฆ๐ด ๐ช๐ฏ๐ท๐ฐ๐ญ๐ท๐ฆ๐ฅ ๐ช๐ฏ ๐ข๐จ๐ฆ๐ฏ๐ต๐ช๐ค ๐ค๐ฐ๐ฎ๐ฎ๐ฆ๐ณ๐ค๐ฆ ๐ด๐ฉ๐ฐ๐ถ๐ญ๐ฅ ๐ฃ๐ฆ ๐ข๐ธ๐ข๐ณ๐ฆ ๐ธ๐ฉ๐ฆ๐ฏ ๐ข๐ฏ ๐๐ ๐ข๐จ๐ฆ๐ฏ๐ต ๐ช๐ด ๐ฉ๐ฆ๐ญ๐ฑ๐ช๐ฏ๐จ ๐ต๐ฐ ๐ฎ๐ข๐ฌ๐ฆ ๐ฐ๐ณ ๐ง๐ข๐ค๐ช๐ญ๐ช๐ต๐ข๐ต๐ฆ ๐ข ๐ต๐ณ๐ข๐ฏ๐ด๐ข๐ค๐ต๐ช๐ฐ๐ฏ, ๐ข๐ฏ๐ฅ ๐ฐ๐ฏ ๐ธ๐ฉ๐ฐ๐ด๐ฆ ๐ฃ๐ฆ๐ฉ๐ข๐ญ๐ง ๐ต๐ฉ๐ฆ๐บ ๐ข๐ณ๐ฆ ๐ข๐ค๐ต๐ช๐ฏ๐จ.”
It’s still up in the air as to whether those building agents will adopt the seemingly obvious principle of telling services that they’re a software agent acting for a human. Do you know an agent is an agent? And if not, ๐ค๐ถ๐ช ๐ฃ๐ฐ๐ฏ๐ฐ?
I’ve built a tool called WhenIsBins for finding out UK bin collection dates. I’ve done so in order to to help me understand agent-related questions like this.
Over the weekend I asked a friend in the US to use Meta’s new Muse personal AI agent to find his UK-based brother’s bin collection dates. Muse had the choice between using the WhenIsBins website, the WhenIsBins API and the supposedly agent-friendly WhenIsBins MCP server.
It chose to use the web version and navigated it as if it were a real person. It did not declare that it was an agent. Others who have looked more deeply suggest that Muse goes out of its way to hide the fact it’s an agent visiting your service. I can’t access Muse in the UK yet, but if that the case that’s… that’s not ideal.
This lack of transparency is the public reason Amazon gave for blocking Muse. And there are lots of reasons why Muse might behave like this, liability being right up there – or rather the avoidance thereof.
But I hope we don’t see an AI-powered arms race between agents pretending to be people and services trying to block agents pretending to be people.
That’s a conflict in which citizens and consumers are guaranteed to end up as losers.
Leave a comment