Re: LLMs and Agents usage in the CCG

You wrote:

> "The core idea: the deploying organization
> registers and signs for its agents, creating a verifiable
> accountability chain that traces back to a known human."

I think this misses the point. *Mailing lists should be for people, not for
machines or "agents." *

A person may use a machine or agent to assist in developing, verifying, or
clarifying ideas, but the active participants in a mailing list should, for
the foreseeable future, be actual humans who actually "own" the words
distributed to others. Claiming accountability for an agent's output
doesn't ensure that a human actually reviews and assumes ownership of the
machine's output before it is broadcast to a list. A useful mailing list
requires ownership, not merely "accountability" for an agent's potential
mistakes or excellent contributions. Merely claiming accountability doesn't
address the issue any more than one's ability to sue the developer of a
"fully-automated vehicle." While knowing who to hold responsible is useful
after an accident, it would be preferable to ensure that machines cannot
autonomously act dangerously.

bob wyman


On Mon, Apr 13, 2026 at 9:18 AM Christian Hommrich <
christian.hommrich@gmail.com> wrote:

> Daniel's point about delegation and credentials for AI members is the
> crux of this.
>
> We've been working on did:trail
> (https://github.com/trailprotocol/trail-did-method) — a W3C DID method
> for AI agent identity. The core idea: the deploying organization
> registers and signs for its agents, creating a verifiable
> accountability chain that traces back to a known human.
>
> Yesterday's Anthropic Managed Agents launch made the gap concrete:
> platform-hosted agents are dynamically provisioned per session — no
> persistent identity in the classical sense. We posted a spec extension
> proposal today addressing this directly:
> https://github.com/trailprotocol/trail-did-method/discussions/10
>
> The accountability model is the same whether the agent is on
> Anthropic, Azure, or self-hosted. The deployer is always accountable.
> The credential is verifiable without platform cooperation.
>
> Happy to discuss whether this fits what the CCG is looking for.
>
> Christian Hommrich
> TRAIL Protocol Initiative
> https://trailprotocol.org
>
>
>