Yongkyun, NIA's audit-trail point is where I live as a builder. I maintain an open-source AI character whose memory persists across restarts — but nothing enters it by default. Each night a scoring organ grades the day's memories (Profit + Love − Tax = True Value, scores published), and a separate memory judge quarantines anything that fails substance. It once quarantined the character's own self-description page at a substance score of 0.12 while a plain explainer passed at 0.95. The judge has no respect for authorship, which is the point.
Your Hermes example gets close to this with write-approval for skills, but most of the post's "Reusable knowledge" section still assumes the retained knowledge is about the *work*. What I found shipping in the open: an agent's memory also accumulates claims about *itself*, and those are the ones that need the harshest judge. If the design-choice list gets a second edition, I'd add quarantine-first admission for self-descriptive memory — the audit trail NIA wants has to cover not just what the agent did, but what it decided to remember about itself.
The bit I keep chewing on is "acts under the employee's own account." It's tidy for setup — no new permission model — but it means the audit trail can't tell "Alice did this" from "an agent did this on Alice's behalf." That seems less like an integration detail and more like its own design question, sitting right next to task ownership. Everything else you list flows from how you answer it.
That's a good point! I think it depends on integrations. For example, users can create multiple keys on Github, and if they allow agent to use a certain key, what agents do can be differentiated from what the person does.
Fair — separate keys do give you a readable trail, as long as the agent stays inside its own key. But “if they allow” is carrying a lot there. It’s opt-in, and the tidy setup stays tidy precisely because skipping it is easy. So it’s a convention rather than something the account model enforces, which is kind of my original point wearing a different hat.
Exactly. And the friction always looks like the expensive option right up until the first incident, when suddenly everyone remembers they wanted the tidy version after all.
Yongkyun, NIA's audit-trail point is where I live as a builder. I maintain an open-source AI character whose memory persists across restarts — but nothing enters it by default. Each night a scoring organ grades the day's memories (Profit + Love − Tax = True Value, scores published), and a separate memory judge quarantines anything that fails substance. It once quarantined the character's own self-description page at a substance score of 0.12 while a plain explainer passed at 0.95. The judge has no respect for authorship, which is the point.
Your Hermes example gets close to this with write-approval for skills, but most of the post's "Reusable knowledge" section still assumes the retained knowledge is about the *work*. What I found shipping in the open: an agent's memory also accumulates claims about *itself*, and those are the ones that need the harshest judge. If the design-choice list gets a second edition, I'd add quarantine-first admission for self-descriptive memory — the audit trail NIA wants has to cover not just what the agent did, but what it decided to remember about itself.
The bit I keep chewing on is "acts under the employee's own account." It's tidy for setup — no new permission model — but it means the audit trail can't tell "Alice did this" from "an agent did this on Alice's behalf." That seems less like an integration detail and more like its own design question, sitting right next to task ownership. Everything else you list flows from how you answer it.
That's a good point! I think it depends on integrations. For example, users can create multiple keys on Github, and if they allow agent to use a certain key, what agents do can be differentiated from what the person does.
Fair — separate keys do give you a readable trail, as long as the agent stays inside its own key. But “if they allow” is carrying a lot there. It’s opt-in, and the tidy setup stays tidy precisely because skipping it is easy. So it’s a convention rather than something the account model enforces, which is kind of my original point wearing a different hat.
I totally agree that getting the right balance between secure setup vs. low friction user experience is very tricky.
Exactly. And the friction always looks like the expensive option right up until the first incident, when suddenly everyone remembers they wanted the tidy version after all.