Assign role boundaries and inspectable handoffs

Verifiedon 2026-07-302026.7.1-2

Isolation is a design decision

OpenClaw documents a multi-agent setup as isolated agents with their own workspace, state directory, and session history, connected by routing bindings. A workspace is a default working directory, not automatically a hard sandbox: absolute host paths remain reachable unless sandboxing is configured. Treat that distinction as a production control, not a detail.

For the capstone, create four role cards:

No role receives a production publishing credential. The publication owner performs the consequential action separately after reviewing the pack.

Lab: make a handoff card

For each edge in your graph, create a card with: task ID; producer role; consumer role; input fingerprint; output location; acceptance criteria; sensitivity label; expiry/deadline; retry count; and stop/rollback action. Add a link or path to the actual artifact—not a private chat transcript.

Validation check: Give the verifier only the researcher card and output. If they cannot reproduce the accept/reject decision, the handoff is incomplete.

Anti-pattern: shared “ops” agent

A single agent with every workspace, every channel binding, and authority to publish may be convenient, but it hides separation of duties and makes incident containment expensive. Start with distinct least-privilege roles; combine them only when the evidence shows the overhead is unjustified and the authority remains explicit.

Learner artifact

Create a redacted, annotated mock of the agents list and bindings view. Highlight one isolated workspace per role, a read-only researcher route, and the absence of a publication-channel binding for all automation roles. Do not use a real account or show credentials.