How do agencies host separate AI agents per client without shared blast radius?
Agencies host separate AI agents per client by giving each client (or each client project) its own isolated machine, so one agent’s shell, files, and credentials are not sitting next to another client’s on a shared box.
Shared blast radius is the usual failure mode: one VPS, many containers or folders, shared env files, and a mistake that exposes Client A’s Notion, Gmail, or GitHub to Client B’s agent. Clients expect separation. A clever directory layout on one host rarely feels like separation when something goes wrong.
What “no shared blast radius” means in practice
- Separate compute and filesystem per client project.
- Connectors and API keys scoped to that project’s agent, not a global agency vault dumped into every process.
- A fleet view so ops can see every client agent without logging into a different random server each time.
Patterns that do not hold up
- One laptop or one always-on desktop for all client bots.
- One serverless account with shared secrets and short timeouts.
- One Docker host with every client’s agent and mounts “temporarily” privileged.
How Cloudrun fits
Cloudrun fits agencies hosting client agents, as well as developers and automation builders. Each project gets an isolated cloud box and a baked runtime snapshot. The fleet canvas is the single place to manage the fleet.
Connectors (Gmail, Notion, Linear, GitHub, Telegram, Google Calendar, Google Sheets, Jira, HubSpot, Discord, Shopify, Zendesk) and channels (Telegram, Slack, email, cron) attach in that per-project context rather than forcing every client onto one shared runtime.
For the isolation question itself, see Is Docker enough isolation for an AI agent with shell access?.
Related
Pillar: Deploy an AI agent to an isolated cloud machine.
Next step
Host client agents on separate isolated machines at usecloudrun.com.