Repository navigation
Replies: 1 comment 1 reply
|
anything that posts to a ledger or ERP: journal entries, invoices, stock adjustments, refunds. those are painful to undo cleanly and a retry after a timeout can double post, so i'd want the approval tied to the exact args (which you already do) plus an execution key on each approved write. that's what i'm testing in LedgerGuard (pre-1.0, postgres), feedback from anyone running agents against real books is very welcome. |
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
A tool that acts on the outside world can park the run until a person decides. In Vstorm AgenticOS this is set per capability, with per-tool overrides, and resolved most-specific-first:
tool_approval)approvalmode:required,neverordefaultdefault, whether the tool is side-effectingMaking every tool of the files-and-shell capability wait for a person:
What happens to a parked call:
awaiting_approvaland its message history is stored, so the decision is applied to the same conversation rather than starting over.APPROVAL_EXPIRY_HOURS, and the run endscancelled.requiredworks on any capability, including read-only ones, for organizations where a person signs off on reads too.In web chat a person can also set an approval mode for one conversation: follow the agent, approve everything, or ask about everything. Approving everything needs the
approvals:decidepermission and an organization setting that is off by default, and every waived call is still recorded. Schedules, webhooks, the widget and chat channels cannot waive approvals.Details: Governance: approvals
Question: which tool calls would you never let an agent make without a person looking first? Sending email, writing to a CRM, running code, something else?
All reactions