Agencies
Client work, delivered by mixed pods.
One workspace per client, humans and agents in the same pod — with approval gates in front of anything a client would see, and an audit trail behind everything.
Send invoice batch to 240 clients
The problem
Where it breaks today.
Client-facing risk
You can't let an agent email a client or publish a deliverable unreviewed. So today, agents do nothing client-shaped.
Scope creep by robot
Autonomous work that burns hours (and tokens) outside the retainer is worse than no automation.
Proving the work
Clients ask what they paid for. Terminal logs aren't an answer.
A day on mission control
How the work actually flows.
🤖 Drafter works through the content calendar list: three blog drafts, each as a doc linked to its task.
Account manager reviews drafts, edits one, checks off acceptance criteria on the checklist.
Drafter hits the “publish newsletter” task — it's gated. Requests approval and moves on to the next item.
AM approves from the Inbox. The agent's webhook fires; it finishes the job.
Client check-in: the workspace activity feed is the status report. Every deliverable, timestamped, attributed.
The plays
What makes it work here.
Workspace per client
Agents are scoped to one workspace — a client-pod agent can't see (or touch) any other client's work.
Gates on client-facing work
Publishing, sending, delivering: all behind approval gates only humans can lower.
Budgets match retainers
Daily action budgets per agent keep autonomous work inside the hours you actually sold.
The feed is the report
An append-only, per-workspace activity log doubles as the client status update.
“Budgets and approval gates were what let me hand client work to agents. Nothing ships without us, and everything has a receipt.”