Trust

Built so someone is always accountable

Agents can't be held accountable. People can. Everything on this page exists to make that workable at the speed agents move.

14:02 maya approved deploy 14:47 sam took over run #142 15:13 agent paused: needs kelly
The model

Many eyes. One owner.

Shared visibility without shared blur. Everyone at the table can see every run; exactly one person owns each run at any moment. Ownership is explicit, visible, and transferable, never ambient.

One owner per run

Agents are delegated to, never assigned. A named person owns every run from start to finish, and the current owner is always visible to the whole table.

Every action has a name on it

Redirects, approvals, hand-offs, stops: recorded with a person and a timestamp. "The agent did it" is never the end of the audit trail here.

Hand-offs are accepted, not assumed

Ownership moves when the incoming person accepts it, and the log records both sides. No orphaned runs, no silent transfers, no "I thought you had it."

The controls

Permission is a noun your team can see

Scoped roles

Each teammate holds specific permissions: deploys, payments, customer data, external sends. Who holds what is visible to everyone, including the agent.

Approvals in the flow

When a run reaches a decision it can't own, it rings the person who holds that permission. The approval, and who gave it, become part of the run's record.

Reject is not undo

Declining agent work never touches a human's undo history and never silently reverts a teammate's work. Agent changes and human changes are separate, always.

Honest state, honest record

Presence means progress, not connection. A run that is burning time without advancing shows as stalled, because a green dot that lies is worse than no dot at all. And the full activity record, every action and every intervention, is exportable, so your source of truth doesn't live only in our product.

Where we are, plainly. Handbell is in early access, built with founding teams. We are not yet audited against SOC 2 or ISO 27001, and we won't imply otherwise; formal certification is planned as the product hardens. What we will do today: answer any security question directly at hello@handbell.ai, share our current architecture and data-handling practices with founding teams, and never train on your data or sell it. If your bar is a completed audit, we'd rather earn your business later than blur that line now.

questions to ask any multiplayer-AI vendor

1  who owns the run, right now, by name?
2  when two people steer at once, who wins?
3  can rejecting agent work undo a human's?
4  is the audit trail attributed and exportable?
5  does a stalled run look stalled?

We publish our answers on this page. Ask everyone else too.

Join the early access