Accord

The more useful my agents become, the more often they reach a point where an action needs my approval. Writing a draft is one thing. Publishing it, changing an account or sending something to another person is different.

A yes in chat can leave too much open. Which version of the message did I approve? Where was it going? If the agent changes the request while I am away, does my earlier answer still apply?

Accord is a Hermes project I'm developing around those questions.

Review the actual request

The agent prepares an action, and Accord captures the tool call, its arguments and its target as a pending request. I can review that request later instead of keeping the conversation open while the agent waits.

The owner can approve, ask for changes, deny or cancel. An approval covers one use of that exact request. If the details change, the changed action needs another review. A comment or a denial doesn't authorize it to run.[6]

A publishing workflow makes the idea concrete. I would want to see the final text and destination, not approve a loose instruction to post something about a project. If I ask for a revision, the revised text should come back before it goes out.

What I'm working toward

I want an agent to prepare useful work while I am busy and bring back a decision I can actually make. That means preserving the request, making the choices clear and returning the result to the right conversation.

Accord is still under development, and this project does not yet authorize live provider actions. Its public source is under the Hermes Human Gate repository name. It also isn't an operating system sandbox: a stronger security boundary needs to keep other write paths and credentials out of the agent's reach.[6]

The part I care about is making approval a real step in the workflow. I should be able to look at what is about to happen, decide, and later check what that decision allowed.