AI Agent Hosting for Agency Handoffs
List the owning client for each synthetic file, the reader observed for each file, and approved cross-client exceptions.
Choose what persists between client jobs
Anthropic documents a self-hosted Claude Agent SDK pattern that creates a container for one user task and destroys it on completion. Consider that pattern for one-off extraction jobs when your agency can own deployment, monitoring, and artifact storage. Use a persistent per-client workspace when the workflow repeatedly reads its own working files, such as an ongoing research brief that carries forward a source list and editorial decisions.
Test the lifecycle you are selling. For a one-off job, save the reviewed deliverable in client-controlled storage, let the job container terminate, then verify that the client can still retrieve the output. For ongoing work, restore the workspace from a backup into a replacement environment and rerun a known task. A chat reply is insufficient if the operator cannot find the files or resume the work within the time promised to the client.
Create synthetic client fixtures with different owners and inspect which trial identity read each one. A successful own-client read is only half of an isolation test.
Hosting the Agent SDK - Claude Code DocsCompare isolation at the boundary the client buys
E2B Enterprise documents a Firecracker microVM per agent session. Consider it when a client requires a session-level execution boundary and your team plans to supply the client portal, approval flow, and durable records. Compare those answers against the client's contract and test the ordinary access paths using two synthetic accounts. If an illustrative client requires execution in its own cloud account, ask E2B about its documented bring-your-own-cloud option for the client's provider and region. Confirm which data and services still use a vendor-operated control plane. Score runtime location, log storage, connector ownership, and offboarding separately. A session isolation feature answers the execution-boundary question. The handoff still needs to prove that the client can retrieve its records and revoke its credentials.
Require an explicit reason for each shared fixture and keep the exception list separate from observed reads. Otherwise a cross-client read can become its own justification.
Enterprise | E2B — The AI Agent Cloud