Agents for your clients, with a trail you can hand over

01

Why raw autonomy is an agency-killing risk

When you run agents on a client’s behalf, every failure is a reputational one you absorb: the outreach email that went to the wrong list, the record update nobody sanctioned, the credential that leaked from a config file. The 2026 OpenClaw incidents — an RCE CVE, tens of thousands of exposed instances, malicious marketplace skills — were mostly individual operators’ problems; for an agency the same failure is multiplied across a client roster and a contract.

Client work also needs boundaries that ad-hoc setups do not have: this client’s work must not bleed into that client’s context, this client’s credentials must not be readable by anyone — including your own staff and the model — and each client’s risk tolerance needs its own policy, not a shared default.

• Per-client sub-user keys (customer_live_) keep each client’s access separately scoped and revocable
• Client credentials go into a write-only vault — stored once, never readable again by anyone
• The autonomy threshold is set per project, so each client gets its own risk posture

02

Running client work through the gate

The delivery loop is concrete: the agent plans a client’s campaign or research brief, drafts the outreach and documents, and files each outward step as a pending proposal; you (or the client, on their own key) review and approve; only then is it dispatched, and the dispatch is logged. Scheduled runs keep retained accounts moving — weekly research digests, pipeline hygiene, drafted follow-ups — without an account manager re-prompting anything. Whatever tools the client already uses connect through MCP; unknown tools fail closed until you decide they are in scope.

The audit trail is what turns this from tooling into a service you can invoice: every proposal, approval, rejection, and dispatch on the account is logged, checkpointed with Ed25519 signatures, and attributable to a key. When the client asks what the agent did in March, you hand over the record. And with the EU AI Act’s Article 14 human-oversight obligations enforceable since August 2026, "our agents propose, a human approves, everything is signed" is a sentence your clients increasingly need to be able to say — the gate lets them say it truthfully.