Which QM Workflows Need a Shared Workspace?

Keep QM for workflows that require coworkers to steer shared work. Pilot managed per-user workspaces for workflows that can move through an authorized handoff with an explicit owner and a verifiable artifact.

Measure what a transfer needs to carry

QM supports personal and shared scopes, and its documentation describes continuity between Slack and the web app. Inventory the workflows that rely on that continuity before deciding what to move. Sample three types of work: a private draft, an asynchronous transfer, and a task where a second person changes direction before the first person finishes. For each sample, write down the minimum state the next operator needs: the artifact revision, outstanding decision, and authority to act. Run the same synthetic case in QM and the candidate workspace, starting the timer when the recipient accepts ownership and stopping when they produce the agreed next artifact. If both participants must steer the same active task, verify that behavior explicitly. A quick export can preserve a draft while leaving a live collaboration requirement unsatisfied.

GitHub - yc-software/qm: Multiplayer agent harness for teams · GitHub

Check the authorization boundary outside the session

Anthropic Managed Agents describes a fresh cloud container per session and a self-hosted option where your worker executes file and shell tool calls. Either model still needs an application-level rule that maps a user to the work they can retrieve. In a pilot, give two test users separate synthetic work products and grant the second user access to only one revision through the approved transfer flow. Verify that the granted revision opens for the recipient, that the private work product stays private through the same application interface, and that the successor records the correct revision. Repeat after starting a fresh session to expose a routing rule that relied on in-memory state. A passing result provides evidence for the tested transfer path. The same rule must be checked wherever your application exposes workspace files.

Choosing an agent environment - Claude Platform Docs

Replay a handoff before picking a host

QM documents personal and shared scopes, with identity and configuration carrying between Slack and its web app. Pick one real handoff that exercises both scopes. Specify who should see the source material, who should be able to redirect the task, and which output may be shared. Score candidates on those actions with separate user accounts.

For an illustrative pilot, suppose an analyst prepares a draft from a private spreadsheet, then a reviewer asks for a revision using only a sanitized summary. Start with sample data and two named test users. Pass the pilot if the reviewer can make the required intervention, both users can find the final output, and the analyst's private input stays private. If the reviewer must steer the original run, test that action before accepting a document-only handoff.

GitHub - yc-software/qm: Multiplayer agent harness for teams · GitHub

Test the direct-agent option on solo work

Trial Claude Code and OpenAI Codex directly when the handoff exercise shows that coworkers only need the reviewed output. Give each tool the same small repository change and require a passing test plus a diff another person can review. Check each vendor's documentation for the account entitlement, execution location, and session behavior you need before counting on any of them. Measure setup time, operator interventions, and whether a second person can review the result from the team's normal review surface.

If the direct route still needs a browser frontend, consider a separate Hermes trial. Hermes documents an OpenAI-compatible HTTP endpoint for frontends. Test the frontend with two identities and the same handoff, including who can retrieve the output and continue the work. A successful API response establishes connectivity. The handoff test establishes whether the collaboration workflow works.

API Server | Hermes Agent

Test isolation and price the support work

If you evaluate a self-operated Hermes workspace, its dashboard documentation says profiles under one OS user can read each other's files. Use separate OS users or containers for tenants who need file separation. Retest access using harmless per-user markers after an upgrade or restore. Treat a cross-user read as a failed migration even when the dashboard shows the expected account.

For an illustrative monthly budget, assume 12 workspaces at a hypothetical $8 each and two hours of operator time at $60 per hour. That is $216 before model usage. Use the same calculation for a self-hosted QM pilot, a managed workspace pilot, and direct agents. Confirm what the vendor quotes per instance or seat, verify disk and usage charges, and include the measured time to provision, recover, and revoke access.

Hermes Web Dashboard | Hermes Agent

Sources